Initial commit for SeaDiff project code

conf.yml

@@ -0,0 +1,54 @@
+# model
+IMAGE_SIZE : [336, 336]         # load image size, if it's train mode, it will be randomly cropped to IMAGE_SIZE. If it's test mode, it will be resized to IMAGE_SIZE.
+CHANNEL_X : 3                   # input channel
+CHANNEL_Y : 3                   # output channel
+TIMESTEPS : 1000                # diffusion steps
+SCHEDULE : 'linear'             # linear or cosine
+MODEL_CHANNELS : 32             # basic channels of Unet
+NUM_RESBLOCKS : 1               # number of residual blocks
+CHANNEL_MULT : [1,2,3,4]        # channel multiplier of each layer
+NUM_HEADS : 1
+DPT_PRETRAINED_WEIGHT: ' '      # path of the pretrained weight of Depth Anything model
+MODE : 1                        # 1 Train, 0 Test
+PRE_ORI : 'True'                # if True, predict $x_0$, else predict $\epsilon$.
+
+# train
+PATH_IMG : ' '                  # path of input
+PATH_GT : ' '                   # path of ground truth
+PATH_GT_DEPTH : ' '             # path of depth
+PATH_IMG_HIST: ' '              # path of histogram
+BATCH_SIZE : 3                  # training batch size
+NUM_WORKERS : 0                 # number of workers
+ITERATION_MAX : 400000          # max training iteration
+LR : 0.0001                     # learning rate
+LOSS : 'L2'                     # L1 or L2
+EMA_EVERY : 100                 # update EMA every EMA_EVERY iterations
+START_EMA : 2000                # start EMA after START_EMA iterations
+SAVE_MODEL_EVERY : 50000        # save model every SAVE_MODEL_EVERY iterations
+EMA: 'False'                    # if True, use EMA
+CONTINUE_TRAINING : 'False'     # if True, continue training
+CONTINUE_TRAINING_STEPS :       # continue training from CONTINUE_TRAINING_STEPS
+
+PRETRAINED_PATH_BETA_PREDICTOR: ' '
+PRETRAINED_PATH_DEPTH_ESTIMATOR : ' '            
+PRETRAINED_PATH_DENOISER: ' '
+
+WEIGHT_SAVE_PATH : ' '          # path to save model
+TRAIN_PATH : ' '                # path of training data
+BETA_LOSS : 50                  # hyperparameter to balance the pixel loss and the diffusion loss
+HIGH_LOW_FREQ : 'False'         # if True, training with frequency separation
+OUTPUT_DIR: ' '
+
+# test
+NATIVE_RESOLUTION : 'False'      # if True, test with native resolution
+DPM_SOLVER : 'False'             # if True, test with DPM_solver
+DPM_STEP : 20                    # DPM_solver step
+BATCH_SIZE_VAL : 1               # test batch size
+PATH_TEST_IMG : ' '              # path of input
+PATH_TEST_GT : ' '               # path of ground truth
+PATH_TEST_GT_DEPTH : ' '         # path of depth
+PATH_TEST_IMG_HIST: ' '          # path of histogram
+
+TEST_PATH : ' '                  # path to save results
+VIS_PATH: ' '                    # path to save results
+

data/data.py

@@ -0,0 +1,89 @@
+import os
+
+import torch
+from PIL import Image
+from torch.utils.data import Dataset
+from torchvision.transforms import (
+    Compose,
+    InterpolationMode,
+    Resize,
+    ToTensor,
+)
+
+
+def ImageTransform(loadSize):
+    return {
+        "train": Compose(
+            [
+                Resize(loadSize, interpolation=InterpolationMode.BILINEAR),
+                ToTensor(),
+            ]
+        ),
+        "test": Compose(
+            [
+                Resize(loadSize, interpolation=InterpolationMode.BILINEAR),
+                ToTensor(),
+            ]
+        ),
+    }
+
+
+class UIEData(Dataset):
+    def __init__(
+        self, path_img, path_gt, path_gt_depth, path_img_hist, loadSize, mode=1
+    ):
+        super().__init__()
+        self.path_img = path_img
+        self.path_gt = path_gt
+        self.path_gt_depth = path_gt_depth
+        self.path_img_hist = path_img_hist
+
+        self.loadsize = loadSize  # e.g. (336, 336) or 336
+        self.crop_pad_size = (
+            loadSize[0] if isinstance(loadSize, (tuple, list)) else loadSize
+        )
+        self.mode = mode
+
+        self.data_img = os.listdir(self.path_img)
+        self.data_gt = os.listdir(self.path_gt)
+        self.data_gt_depth = os.listdir(self.path_gt_depth)
+        self.data_img_hist = os.listdir(self.path_img_hist)
+        if mode == 1:
+            self.ImgTrans = ImageTransform(loadSize)["train"]
+        else:
+            self.ImgTrans = ImageTransform(loadSize)["test"]
+
+    def __len__(self):
+        return len(self.data_gt)
+
+    def __getitem__(self, idx):
+        img = Image.open(os.path.join(self.path_img, self.data_img[idx])).convert("RGB")
+        gt = Image.open(os.path.join(self.path_gt, self.data_img[idx])).convert("RGB")
+        label_depth = Image.open(
+            os.path.join(self.path_gt_depth, self.data_img[idx])
+        ).convert("RGB")
+        img_hist = Image.open(
+            os.path.join(self.path_img_hist, self.data_img[idx])
+        ).convert("RGB")
+
+        name = self.data_img[idx]
+        h, w = img.size
+
+        if self.mode == 1:
+            seed = torch.random.seed()
+            torch.random.manual_seed(seed)
+            img = self.ImgTrans(img)
+            torch.random.manual_seed(seed)
+            gt = self.ImgTrans(gt)
+            torch.random.manual_seed(seed)
+            label_depth = self.ImgTrans(label_depth)
+            torch.random.manual_seed(seed)
+            img_hist = self.ImgTrans(img_hist)
+
+        else:
+            img = self.ImgTrans(img)
+            gt = self.ImgTrans(gt)
+            label_depth = self.ImgTrans(label_depth)
+            img_hist = self.ImgTrans(img_hist)
+
+        return img, gt, label_depth, img_hist, name, (h, w)

depth_anything/blocks.py

@@ -0,0 +1,153 @@
+import torch.nn as nn
+
+
+def _make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    if len(in_shape) >= 4:
+        out_shape4 = out_shape
+
+    if expand:
+        out_shape1 = out_shape
+        out_shape2 = out_shape*2
+        out_shape3 = out_shape*4
+        if len(in_shape) >= 4:
+            out_shape4 = out_shape*8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    if len(in_shape) >= 4:
+        scratch.layer4_rn = nn.Conv2d(
+            in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+        )
+
+    return scratch
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features, activation, bn):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups=1
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+        
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+
+        if self.bn==True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn==True:
+            out = self.bn1(out)
+       
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn==True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, size=None):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups=1
+
+        self.expand = expand
+        out_features = features
+        if self.expand==True:
+            out_features = features//2
+        
+        self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
+
+        self.resConfUnit1 = ResidualConvUnit(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit(features, activation, bn)
+        
+        self.skip_add = nn.quantized.FloatFunctional()
+
+        self.size=size
+
+    def forward(self, *xs, size=None):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+
+        output = self.resConfUnit2(output)
+
+        if (size is None) and (self.size is None):
+            modifier = {"scale_factor": 2}
+        elif size is None:
+            modifier = {"size": self.size}
+        else:
+            modifier = {"size": size}
+
+        output = nn.functional.interpolate(
+            output, **modifier, mode="bilinear", align_corners=self.align_corners
+        )
+
+        output = self.out_conv(output)
+
+        return output

depth_anything/dpt.py

@@ -0,0 +1,171 @@
+import torch
+import torch.nn as nn
+
+from .blocks import FeatureFusionBlock, _make_scratch
+import torch.nn.functional as F
+
+
+def _make_fusion_block(features, use_bn, size = None):
+    return FeatureFusionBlock(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+        size=size,
+    )
+
+
+class DPTHead(nn.Module):
+    def __init__(self, nclass, in_channels, features=256, use_bn=False, out_channels=[256, 512, 1024, 1024], use_clstoken=False):
+        super(DPTHead, self).__init__()
+        
+        self.nclass = nclass
+        self.use_clstoken = use_clstoken
+        
+        self.projects = nn.ModuleList([
+            nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channel,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ) for out_channel in out_channels
+        ])
+        
+        self.resize_layers = nn.ModuleList([
+            nn.ConvTranspose2d(
+                in_channels=out_channels[0],
+                out_channels=out_channels[0],
+                kernel_size=4,
+                stride=4,
+                padding=0),
+            nn.ConvTranspose2d(
+                in_channels=out_channels[1],
+                out_channels=out_channels[1],
+                kernel_size=2,
+                stride=2,
+                padding=0),
+            nn.Identity(),
+            nn.Conv2d(
+                in_channels=out_channels[3],
+                out_channels=out_channels[3],
+                kernel_size=3,
+                stride=2,
+                padding=1)
+        ])
+        
+        if use_clstoken:
+            self.readout_projects = nn.ModuleList()
+            for _ in range(len(self.projects)):
+                self.readout_projects.append(
+                    nn.Sequential(
+                        nn.Linear(2 * in_channels, in_channels),
+                        nn.GELU()))
+        
+        self.scratch = _make_scratch(
+            out_channels,
+            features,
+            groups=1,
+            expand=False,
+        )
+
+        self.scratch.stem_transpose = None
+        
+        self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
+
+        head_features_1 = features
+        head_features_2 = 32
+        
+        if nclass > 1:
+            self.scratch.output_conv = nn.Sequential(
+                nn.Conv2d(head_features_1, head_features_1, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(True),
+                nn.Conv2d(head_features_1, nclass, kernel_size=1, stride=1, padding=0),
+            )
+        else:
+            self.scratch.output_conv1 = nn.Conv2d(head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1)
+            
+            self.scratch.output_conv2 = nn.Sequential(
+                nn.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(True),
+                nn.Conv2d(head_features_2, 1, kernel_size=1, stride=1, padding=0),
+                nn.ReLU(True),
+                nn.Identity(),
+            )
+            
+    def forward(self, out_features, patch_h, patch_w):
+        out = []
+        for i, x in enumerate(out_features):
+            if self.use_clstoken:
+                x, cls_token = x[0], x[1]
+                readout = cls_token.unsqueeze(1).expand_as(x)
+                x = self.readout_projects[i](torch.cat((x, readout), -1))
+            else:
+                x = x[0]
+            
+            x = x.permute(0, 2, 1).reshape((x.shape[0], x.shape[-1], patch_h, patch_w))
+            
+            x = self.projects[i](x)
+            x = self.resize_layers[i](x)
+            
+            out.append(x)
+        
+        layer_1, layer_2, layer_3, layer_4 = out
+        
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+        
+        path_4 = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:])
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn, size=layer_2_rn.shape[2:])
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn, size=layer_1_rn.shape[2:])
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+        
+        out = self.scratch.output_conv1(path_1)
+        out = F.interpolate(out, (int(patch_h * 14), int(patch_w * 14)), mode="bilinear", align_corners=True)
+        out = self.scratch.output_conv2(out)
+        
+        return out
+        
+        
+class DPT_DINOv2(nn.Module):
+    def __init__(self, encoder='vitl', features=256, out_channels=[256, 512, 1024, 1024], use_bn=False, use_clstoken=False, localhub=True):
+        super(DPT_DINOv2, self).__init__()
+        
+        assert encoder in ['vits', 'vitb', 'vitl']
+        
+        # in case the Internet connection is not stable, please load the DINOv2 locally
+        if localhub:
+            self.pretrained = torch.hub.load('torchhub/facebookresearch_dinov2_main', 'dinov2_{:}14'.format(encoder), source='local', pretrained=False)
+        else:
+            self.pretrained = torch.hub.load('facebookresearch/dinov2', 'dinov2_{:}14'.format(encoder))
+        
+        dim = self.pretrained.blocks[0].attn.qkv.in_features
+        
+        self.depth_head = DPTHead(1, dim, features, use_bn, out_channels=out_channels, use_clstoken=use_clstoken)
+        
+    def forward(self, x):
+        h, w = x.shape[-2:]
+        
+        features = self.pretrained.get_intermediate_layers(x, 4, return_class_token=True)
+        
+        patch_h, patch_w = h // 14, w // 14
+
+        depth = self.depth_head(features, patch_h, patch_w)
+        depth = F.interpolate(depth, size=(h, w), mode="bilinear", align_corners=True)
+        depth = F.relu(depth)
+
+        # return depth.squeeze(1)
+        return depth
+
+
+if __name__ == '__main__':
+    depth_anything = DPT_DINOv2()
+    depth_anything.load_state_dict(torch.load('checkpoints/depth_anything_dinov2_vitl14.pth'))
+    

depth_anything/util/transform.py

@@ -0,0 +1,248 @@
+import random
+from PIL import Image, ImageOps, ImageFilter
+import torch
+from torchvision import transforms
+import torch.nn.functional as F
+
+import numpy as np
+import cv2
+import math
+
+
+def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
+    """Rezise the sample to ensure the given size. Keeps aspect ratio.
+
+    Args:
+        sample (dict): sample
+        size (tuple): image size
+
+    Returns:
+        tuple: new size
+    """
+    shape = list(sample["disparity"].shape)
+
+    if shape[0] >= size[0] and shape[1] >= size[1]:
+        return sample
+
+    scale = [0, 0]
+    scale[0] = size[0] / shape[0]
+    scale[1] = size[1] / shape[1]
+
+    scale = max(scale)
+
+    shape[0] = math.ceil(scale * shape[0])
+    shape[1] = math.ceil(scale * shape[1])
+
+    # resize
+    sample["image"] = cv2.resize(
+        sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
+    )
+
+    sample["disparity"] = cv2.resize(
+        sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
+    )
+    sample["mask"] = cv2.resize(
+        sample["mask"].astype(np.float32),
+        tuple(shape[::-1]),
+        interpolation=cv2.INTER_NEAREST,
+    )
+    sample["mask"] = sample["mask"].astype(bool)
+
+    return tuple(shape)
+
+
+class Resize(object):
+    """Resize sample to given size (width, height).
+    """
+
+    def __init__(
+        self,
+        width,
+        height,
+        resize_target=True,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resize_method="lower_bound",
+        image_interpolation_method=cv2.INTER_AREA,
+    ):
+        """Init.
+
+        Args:
+            width (int): desired output width
+            height (int): desired output height
+            resize_target (bool, optional):
+                True: Resize the full sample (image, mask, target).
+                False: Resize image only.
+                Defaults to True.
+            keep_aspect_ratio (bool, optional):
+                True: Keep the aspect ratio of the input sample.
+                Output sample might not have the given width and height, and
+                resize behaviour depends on the parameter 'resize_method'.
+                Defaults to False.
+            ensure_multiple_of (int, optional):
+                Output width and height is constrained to be multiple of this parameter.
+                Defaults to 1.
+            resize_method (str, optional):
+                "lower_bound": Output will be at least as large as the given size.
+                "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
+                "minimal": Scale as least as possible.  (Output size might be smaller than given size.)
+                Defaults to "lower_bound".
+        """
+        self.__width = width
+        self.__height = height
+
+        self.__resize_target = resize_target
+        self.__keep_aspect_ratio = keep_aspect_ratio
+        self.__multiple_of = ensure_multiple_of
+        self.__resize_method = resize_method
+        self.__image_interpolation_method = image_interpolation_method
+
+    def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
+        y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        return y
+
+    def get_size(self, width, height):
+        # determine new height and width
+        scale_height = self.__height / height
+        scale_width = self.__width / width
+
+        if self.__keep_aspect_ratio:
+            if self.__resize_method == "lower_bound":
+                # scale such that output size is lower bound
+                if scale_width > scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "upper_bound":
+                # scale such that output size is upper bound
+                if scale_width < scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "minimal":
+                # scale as least as possbile
+                if abs(1 - scale_width) < abs(1 - scale_height):
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            else:
+                raise ValueError(
+                    f"resize_method {self.__resize_method} not implemented"
+                )
+
+        if self.__resize_method == "lower_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, min_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, min_val=self.__width
+            )
+        elif self.__resize_method == "upper_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, max_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, max_val=self.__width
+            )
+        elif self.__resize_method == "minimal":
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+        else:
+            raise ValueError(f"resize_method {self.__resize_method} not implemented")
+
+        return (new_width, new_height)
+
+    def __call__(self, sample):
+        width, height = self.get_size(
+            sample["image"].shape[1], sample["image"].shape[0]
+        )
+
+        # resize sample
+        sample["image"] = cv2.resize(
+            sample["image"],
+            (width, height),
+            interpolation=self.__image_interpolation_method,
+        )
+
+        if self.__resize_target:
+            if "disparity" in sample:
+                sample["disparity"] = cv2.resize(
+                    sample["disparity"],
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+
+            if "depth" in sample:
+                sample["depth"] = cv2.resize(
+                    sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
+                )
+
+            if "semseg_mask" in sample:
+                # sample["semseg_mask"] = cv2.resize(
+                #     sample["semseg_mask"], (width, height), interpolation=cv2.INTER_NEAREST
+                # )
+                sample["semseg_mask"] = F.interpolate(torch.from_numpy(sample["semseg_mask"]).float()[None, None, ...], (height, width), mode='nearest').numpy()[0, 0]
+                
+            if "mask" in sample:
+                sample["mask"] = cv2.resize(
+                    sample["mask"].astype(np.float32),
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+                # sample["mask"] = sample["mask"].astype(bool)
+
+        # print(sample['image'].shape, sample['depth'].shape)
+        return sample
+
+
+class NormalizeImage(object):
+    """Normlize image by given mean and std.
+    """
+
+    def __init__(self, mean, std):
+        self.__mean = mean
+        self.__std = std
+
+    def __call__(self, sample):
+        sample["image"] = (sample["image"] - self.__mean) / self.__std
+
+        return sample
+
+
+class PrepareForNet(object):
+    """Prepare sample for usage as network input.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, sample):
+        image = np.transpose(sample["image"], (2, 0, 1))
+        sample["image"] = np.ascontiguousarray(image).astype(np.float32)
+
+        if "mask" in sample:
+            sample["mask"] = sample["mask"].astype(np.float32)
+            sample["mask"] = np.ascontiguousarray(sample["mask"])
+        
+        if "depth" in sample:
+            depth = sample["depth"].astype(np.float32)
+            sample["depth"] = np.ascontiguousarray(depth)
+            
+        if "semseg_mask" in sample:
+            sample["semseg_mask"] = sample["semseg_mask"].astype(np.float32)
+            sample["semseg_mask"] = np.ascontiguousarray(sample["semseg_mask"])
+
+        return sample

main.py

@@ -0,0 +1,18 @@
+from src.config import load_config
+from src.train import train, test
+import argparse
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--config', type=str, default='conf.yml', help='path to the config.yaml file')
+    args = parser.parse_args()
+    config = load_config(args.config)
+    print('Config loaded')
+    mode = config.MODE
+    if mode == 1:
+        train(config)
+    else:
+        test(config)
+if __name__ == "__main__":
+    main()

model/DocDiff.py

@@ -0,0 +1,512 @@
+import math
+
+import numpy as np
+import torch
+from torch import nn
+from utils.utils import get_A
+
+
+class Swish(nn.Module):
+    """
+    ### Swish actiavation function
+    $$x \cdot \sigma(x)$$
+    """
+
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class TimeEmbedding(nn.Module):
+    """
+    ### Embeddings for $t$
+    """
+
+    def __init__(self, n_channels: int):
+        """
+        * `n_channels` is the number of dimensions in the embedding
+        """
+        super().__init__()
+        self.n_channels = n_channels
+        # First linear layer
+        self.lin1 = nn.Linear(self.n_channels // 4, self.n_channels)
+        # Activation
+        self.act = Swish()
+        # Second linear layer
+        self.lin2 = nn.Linear(self.n_channels, self.n_channels)
+
+    def forward(self, t: torch.Tensor):
+        # Create sinusoidal position embeddings
+        # [same as those from the transformer](../../transformers/positional_encoding.html)
+        #
+        # \begin{align}
+        # PE^{(1)}_{t,i} &= sin\Bigg(\frac{t}{10000^{\frac{i}{d - 1}}}\Bigg) \\
+        # PE^{(2)}_{t,i} &= cos\Bigg(\frac{t}{10000^{\frac{i}{d - 1}}}\Bigg)
+        # \end{align}
+        #
+        # where $d$ is `half_dim`
+        half_dim = self.n_channels // 8
+        emb = math.log(10_000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
+        emb = t[:, None] * emb[None, :]
+        emb = torch.cat((emb.sin(), emb.cos()), dim=1)
+
+        # Transform with the MLP
+        emb = self.act(self.lin1(emb))
+        emb = self.lin2(emb)
+
+        #
+        return emb
+
+
+class ResidualBlock(nn.Module):
+    """
+    ### Residual block
+    A residual block has two convolution layers with group normalization.
+    Each resolution is processed with two residual blocks.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        time_channels: int,
+        dropout: float = 0.1,
+        is_noise: bool = True,
+    ):
+        """
+        * `in_channels` is the number of input channels
+        * `out_channels` is the number of input channels
+        * `time_channels` is the number channels in the time step ($t$) embeddings
+        * `n_groups` is the number of groups for [group normalization](../../normalization/group_norm/index.html)
+        * `dropout` is the dropout rate
+        """
+        super().__init__()
+        # Group normalization and the first convolution layer
+        self.is_noise = is_noise
+        self.act1 = Swish()
+        self.conv1 = nn.Conv2d(
+            in_channels, out_channels, kernel_size=(3, 3), padding=(1, 1)
+        )
+
+        # Group normalization and the second convolution layer
+
+        self.act2 = Swish()
+        self.conv2 = nn.Conv2d(
+            out_channels, out_channels, kernel_size=(3, 3), padding=(1, 1)
+        )
+
+        # If the number of input channels is not equal to the number of output channels we have to
+        # project the shortcut connection
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=(1, 1))
+        else:
+            self.shortcut = nn.Identity()
+
+        # Linear layer for time embeddings
+        if self.is_noise:
+            self.time_emb = nn.Linear(time_channels, out_channels)
+            self.time_act = Swish()
+
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        """
+        * `x` has shape `[batch_size, in_channels, height, width]`
+        * `t` has shape `[batch_size, time_channels]`
+        """
+        # First convolution layer
+        h = self.conv1(self.act1(x))
+        # Add time embeddings
+        if self.is_noise:
+            h += self.time_emb(self.time_act(t))[:, :, None, None]
+        # Second convolution layer
+        h = self.conv2(self.dropout(self.act2(h)))
+
+        # Add the shortcut connection and return
+        return h + self.shortcut(x)
+
+
+class DownBlock(nn.Module):
+    """
+    ### Down block
+    This combines `ResidualBlock` and `AttentionBlock`. These are used in the first half of U-Net at each resolution.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        time_channels: int,
+        is_noise: bool = True,
+    ):
+        super().__init__()
+        self.res = ResidualBlock(
+            in_channels, out_channels, time_channels, is_noise=is_noise
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        x = self.res(x, t)
+        return x
+
+
+class UpBlock(nn.Module):
+    """
+    ### Up block
+    This combines `ResidualBlock` and `AttentionBlock`. These are used in the second half of U-Net at each resolution.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        time_channels: int,
+        is_noise: bool = True,
+    ):
+        super().__init__()
+        # The input has `in_channels + out_channels` because we concatenate the output of the same resolution
+        # from the first half of the U-Net
+        self.res = ResidualBlock(
+            in_channels + out_channels, out_channels, time_channels, is_noise=is_noise
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        x = self.res(x, t)
+        return x
+
+
+class MiddleBlock(nn.Module):
+    """
+    ### Middle block
+    It combines a `ResidualBlock`, `AttentionBlock`, followed by another `ResidualBlock`.
+    This block is applied at the lowest resolution of the U-Net.
+    """
+
+    def __init__(self, n_channels: int, time_channels: int, is_noise: bool = True):
+        super().__init__()
+        self.res1 = ResidualBlock(
+            n_channels, n_channels, time_channels, is_noise=is_noise
+        )
+        self.dia1 = nn.Conv2d(
+            n_channels, n_channels, 3, 1, dilation=2, padding=get_pad(16, 3, 1, 2)
+        )
+        self.dia2 = nn.Conv2d(
+            n_channels, n_channels, 3, 1, dilation=4, padding=get_pad(16, 3, 1, 4)
+        )
+        self.dia3 = nn.Conv2d(
+            n_channels, n_channels, 3, 1, dilation=8, padding=get_pad(16, 3, 1, 8)
+        )
+        self.dia4 = nn.Conv2d(
+            n_channels, n_channels, 3, 1, dilation=16, padding=get_pad(16, 3, 1, 16)
+        )
+        self.res2 = ResidualBlock(
+            n_channels, n_channels, time_channels, is_noise=is_noise
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        x = self.res1(x, t)
+        x = self.dia1(x)
+        x = self.dia2(x)
+        x = self.dia3(x)
+        x = self.dia4(x)
+        x = self.res2(x, t)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    ### Scale up the feature map by $2 \times$
+    """
+
+    def __init__(self, n_channels):
+        super().__init__()
+        self.conv = nn.ConvTranspose2d(n_channels, n_channels, (4, 4), (2, 2), (1, 1))
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        # `t` is not used, but it's kept in the arguments because for the attention layer function signature
+        # to match with `ResidualBlock`.
+        _ = t
+        return self.conv(x)
+
+
+class Downsample(nn.Module):
+    """
+    ### Scale down the feature map by $\frac{1}{2} \times$
+    """
+
+    def __init__(self, n_channels):
+        super().__init__()
+        self.conv = nn.Conv2d(n_channels, n_channels, (3, 3), (2, 2), (1, 1))
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor):
+        # `t` is not used, but it's kept in the arguments because for the attention layer function signature
+        # to match with `ResidualBlock`.
+        _ = t
+        return self.conv(x)
+
+
+class Denoise_UNet(nn.Module):
+    """
+    ## U-Net
+    """
+
+    def __init__(
+        self, input_channels, output_channels, n_channels, ch_mults, n_blocks, is_noise
+    ):
+        """
+        * `image_channels` is the number of channels in the image. $3$ for RGB.
+        * `n_channels` is number of channels in the initial feature map that we transform the image into
+        * `ch_mults` is the list of channel numbers at each resolution. The number of channels is `ch_mults[i] * n_channels`
+        * `is_attn` is a list of booleans that indicate whether to use attention at each resolution
+        * `n_blocks` is the number of `UpDownBlocks` at each resolution
+        """
+        super().__init__()
+
+        # Number of resolutions
+        n_resolutions = len(ch_mults)
+
+        # Project image into feature map
+        self.image_proj = nn.Conv2d(
+            input_channels, n_channels, kernel_size=(3, 3), padding=(1, 1)
+        )
+
+        # Time embedding layer. Time embedding has `n_channels * 4` channels
+        self.is_noise = is_noise
+        if is_noise:
+            self.time_emb = TimeEmbedding(n_channels * 4)
+
+        # #### First half of U-Net - decreasing resolution
+        down = []
+        # Number of channels
+        out_channels = in_channels = n_channels
+        # For each resolution
+        for i in range(n_resolutions):
+            # Number of output channels at this resolution
+            out_channels = n_channels * ch_mults[i]
+            # Add `n_blocks`
+            for _ in range(n_blocks):
+                down.append(
+                    DownBlock(
+                        in_channels, out_channels, n_channels * 4, is_noise=is_noise
+                    )
+                )
+                in_channels = out_channels
+            # Down sample at all resolutions except the last
+            if i < n_resolutions - 1:
+                down.append(Downsample(in_channels))
+
+        # Combine the set of modules
+        self.down = nn.ModuleList(down)
+
+        # Middle block
+        self.middle = MiddleBlock(out_channels, n_channels * 4, is_noise=False)
+
+        # #### Second half of U-Net - increasing resolution
+        up = []
+        # Number of channels
+        in_channels = out_channels
+        # For each resolution
+        for i in reversed(range(n_resolutions)):
+            # `n_blocks` at the same resolution
+            out_channels = n_channels * ch_mults[i]
+            for _ in range(n_blocks):
+                up.append(
+                    UpBlock(
+                        in_channels, out_channels, n_channels * 4, is_noise=is_noise
+                    )
+                )
+            # Final block to reduce the number of channels
+            in_channels = n_channels * (ch_mults[i - 1] if i >= 1 else 1)
+            up.append(
+                UpBlock(in_channels, out_channels, n_channels * 4, is_noise=is_noise)
+            )
+            in_channels = out_channels
+            # Up sample at all resolutions except last
+            if i > 0:
+                up.append(Upsample(in_channels))
+
+        # Combine the set of modules
+        self.up = nn.ModuleList(up)
+
+        self.act = Swish()
+        self.final = nn.Conv2d(
+            in_channels, output_channels, kernel_size=(3, 3), padding=(1, 1)
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor = torch.tensor([0]).cuda()):
+        """
+        * `x` has shape `[batch_size, in_channels, height, width]`
+        * `t` has shape `[batch_size]`
+        """
+
+        # Get time-step embeddings
+        if self.is_noise:
+            t = self.time_emb(t)
+        else:
+            t = None
+        # Get image projection
+        x = self.image_proj(x)
+
+        # `h` will store outputs at each resolution for skip connection
+        h = [x]
+        # First half of U-Net
+        for m in self.down:
+            x = m(x, t)
+            h.append(x)
+
+        # Middle (bottom)
+        x = self.middle(x, t)
+
+        # Second half of U-Net
+        for m in self.up:
+            if isinstance(m, Upsample):
+                x = m(x, t)
+            else:
+                # Get the skip connection from first half of U-Net and concatenate
+                s = h.pop()
+                # print(x.shape, s.shape)
+                x = torch.cat((x, s), dim=1)
+                #
+                x = m(x, t)
+
+        # Final normalization and convolution
+        return self.final(self.act(x))
+
+
+class Beta_UNet(nn.Module):
+    def __init__(self, input_channels, output_channels, n_channels, ch_mults, n_blocks):
+        super().__init__()
+        is_noise = False
+        # Number of resolutions
+        n_resolutions = len(ch_mults)
+
+        # Project image into feature map
+        self.image_proj = nn.Conv2d(
+            input_channels, n_channels, kernel_size=(3, 3), padding=(1, 1)
+        )
+
+        # #### First half of U-Net - decreasing resolution
+        down = []
+        # Number of channels
+        out_channels = in_channels = n_channels
+        # For each resolution
+        for i in range(n_resolutions):
+            # Number of output channels at this resolution
+            out_channels = n_channels * ch_mults[i]
+            # Add `n_blocks`
+            for _ in range(n_blocks):
+                down.append(
+                    DownBlock(
+                        in_channels, out_channels, n_channels * 4, is_noise=is_noise
+                    )
+                )
+                in_channels = out_channels
+            # Down sample at all resolutions except the last
+            if i < n_resolutions - 1:
+                down.append(Downsample(in_channels))
+        # Combine the set of modules
+        self.down = nn.ModuleList(down)
+
+        # Middle block
+        self.middle = MiddleBlock(out_channels, n_channels * 4, is_noise=False)
+        self.act = Swish()
+
+        self.pool = nn.AdaptiveAvgPool2d((1, 1))
+
+        self.transform = nn.Sequential(nn.Linear(128, 3), Swish(), nn.Linear(3, 3))
+
+    def forward(self, x: torch.Tensor):
+        t = None
+        x = self.image_proj(x)
+        h = [x]
+        for m in self.down:
+            x = m(x, t)
+            h.append(x)
+        x = self.middle(x, t)
+        x = torch.sigmoid(self.transform(self.pool(x).squeeze()))
+        return x.unsqueeze(-1).unsqueeze(-1)
+
+
+class DocDiff(nn.Module):
+    def __init__(
+        self,
+        input_channels,
+        output_channels,
+        n_channels,
+        ch_mults,
+        n_blocks,
+    ):
+        super(DocDiff, self).__init__()
+        self.beta_predictor = Beta_UNet(3, 3, n_channels, ch_mults, n_blocks)
+        self.denoiser = Denoise_UNet(
+            12, 3, n_channels, ch_mults, n_blocks, is_noise=True
+        )
+
+    def forward(self, x, condition, hist, depth, t, diffusion):
+        pred_beta = self.beta_predictor(condition)
+        depth = (depth - depth.min()) / (depth.max() - depth.min())
+        T_direct = torch.clamp((torch.exp(-pred_beta * depth)), 0, 1)
+        T_scatter = torch.clamp((1 - torch.exp(-pred_beta * depth)), 0, 1)
+        atm_light = [get_A(item) for item in condition]
+        atm_light = torch.stack(atm_light).to(x.device)
+        J = torch.clamp(((condition - T_scatter * atm_light) / T_direct), 0, 1)
+
+        noisy_image, noise_ref = diffusion.noisy_image(t, x)
+        denoised_J = self.denoiser(
+            torch.cat((noisy_image, condition.clone().detach(), J, hist), dim=1), t
+        )
+        return J, noise_ref, denoised_J, T_direct, T_scatter
+
+
+class EMA:
+    def __init__(self, beta):
+        super().__init__()
+        self.beta = beta
+
+    def update_model_average(self, ma_model, current_model):
+        for current_params, ma_params in zip(
+            current_model.parameters(), ma_model.parameters()
+        ):
+            old_weight, up_weight = ma_params.data, current_params.data
+            ma_params.data = self.update_average(old_weight, up_weight)
+
+    def update_average(self, old, new):
+        if old is None:
+            return new
+        return old * self.beta + (1 - self.beta) * new
+
+
+def get_pad(in_, ksize, stride, atrous=1):
+    out_ = np.ceil(float(in_) / stride)
+    return int(((out_ - 1) * stride + atrous * (ksize - 1) + 1 - in_) / 2)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    import torchsummary
+    from schedule.diffusionSample import GaussianDiffusion
+    from schedule.schedule import Schedule
+    from src.config import load_config
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--config", type=str, default="../conf.yml", help="path to the config.yaml file"
+    )
+    args = parser.parse_args()
+    config = load_config(args.config)
+    print("Config loaded")
+    model = DocDiff(
+        input_channels=config.CHANNEL_X + config.CHANNEL_Y,
+        output_channels=config.CHANNEL_Y,
+        n_channels=config.MODEL_CHANNELS,
+        ch_mults=config.CHANNEL_MULT,
+        n_blocks=config.NUM_RESBLOCKS,
+    )
+    schedule = Schedule(config.SCHEDULE, config.TIMESTEPS)
+    diffusion = GaussianDiffusion(model, config.TIMESTEPS, schedule)
+    model.eval()
+    print(
+        torchsummary.summary(
+            model.init_predictor.cuda(), [(3, 128, 128)], batch_size=32
+        )
+    )

schedule/diffusionSample.py

@@ -0,0 +1,140 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def extract_(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def extract(v, t, x_shape):
+    """
+    Extract some coefficients at specified timesteps, then reshape to
+    [batch_size, 1, 1, 1, 1, ...] for broadcasting purposes.
+    """
+    device = t.device
+    out = torch.gather(v, index=t, dim=0).float().to(device)
+    return out.view([t.shape[0]] + [1] * (len(x_shape) - 1))
+
+
+class GaussianDiffusion(nn.Module):
+    def __init__(self, model, T, schedule):
+        super().__init__()
+        self.visual = False
+        if self.visual:
+            self.num = 0
+        self.model = model
+        self.T = T
+        self.schedule = schedule
+        betas = self.schedule.get_betas()
+        self.register_buffer("betas", betas.float())
+        alphas = 1.0 - self.betas
+        alphas_bar = torch.cumprod(alphas, dim=0)
+        alphas_bar_prev = F.pad(alphas_bar, [1, 0], value=1)[:T]
+        gammas = alphas_bar
+
+        self.register_buffer("coeff1", torch.sqrt(1.0 / alphas))
+        self.register_buffer(
+            "coeff2", self.coeff1 * (1.0 - alphas) / torch.sqrt(1.0 - alphas_bar)
+        )
+        self.register_buffer(
+            "posterior_var", self.betas * (1.0 - alphas_bar_prev) / (1.0 - alphas_bar)
+        )
+
+        self.register_buffer("gammas", gammas)
+        self.register_buffer("sqrt_one_minus_gammas", torch.sqrt(1 - gammas))
+        self.register_buffer("sqrt_gammas", torch.sqrt(gammas))
+
+    def predict_xt_prev_mean_from_eps(self, x_t, t, eps):
+        assert x_t.shape == eps.shape
+        return (
+            extract(self.coeff1, t, x_t.shape) * x_t
+            - extract(self.coeff2, t, x_t.shape) * eps
+        )
+
+    def predict_eps_from_x0(self, x_t, t, x_0):
+        return (x_t - extract(self.sqrt_gammas, t, x_t.shape) * x_0) / extract(
+            self.sqrt_one_minus_gammas, t, x_t.shape
+        )
+
+    def x0_p_mean_variance(self, x_t, cond_, t):
+        var = torch.cat([self.posterior_var[1:2], self.betas[1:]])
+        var = extract(var, t, x_t.shape)
+
+        x0_pred = self.model(torch.cat((x_t, cond_), dim=1), t)
+
+        eps = self.predict_eps_from_x0(x_t, t, x0_pred)
+
+        xt_prev_mean = self.predict_xt_prev_mean_from_eps(x_t, t, eps=eps)
+
+        return xt_prev_mean, var
+
+    def p_mean_variance(self, x_t, cond_, t):
+        var = torch.cat([self.posterior_var[1:2], self.betas[1:]])
+        var = extract(var, t, x_t.shape)
+
+        eps = self.model(torch.cat((x_t, cond_), dim=1), t)
+
+        xt_prev_mean = self.predict_xt_prev_mean_from_eps(x_t, t, eps=eps)
+
+        return xt_prev_mean, var
+
+    def noisy_image(self, t, y):
+        """Compute y_noisy according to (6) p15 of [2]"""
+        noise = torch.randn_like(y)
+        y_noisy = (
+            extract_(self.sqrt_gammas, t, y.shape) * y
+            + extract_(self.sqrt_one_minus_gammas, t, noise.shape) * noise
+        )
+        return y_noisy, noise
+
+    def forward(self, x_T, cond, cond_J, cond_hist, pre_ori="False"):
+        """
+        Algorithm 2.
+        """
+        x_t = x_T
+        cond_ = cond
+        cond_hist_ = cond_hist
+        cond_J_ = cond_J
+        for time_step in reversed(range(self.T)):
+            print("time_step: ", time_step)
+            t = (
+                x_t.new_ones(
+                    [
+                        x_T.shape[0],
+                    ],
+                    dtype=torch.long,
+                )
+                * time_step
+            )
+            if pre_ori == "False":
+                mean, var = self.p_mean_variance(
+                    x_t=x_t, t=t, cond_=torch.cat((cond_, cond_J_, cond_hist_), dim=1)
+                )
+                if time_step > 0:
+                    noise = torch.randn_like(x_t)
+                else:
+                    noise = 0
+                x_t = mean + torch.sqrt(var) * noise
+                assert torch.isnan(x_t).int().sum() == 0, "nan in tensor."
+            else:
+                mean, var = self.x0_p_mean_variance(
+                    x_t=x_t, t=t, cond_=torch.cat((cond_, cond_J_, cond_hist_), dim=1)
+                )
+                if time_step > 0:
+                    noise = torch.randn_like(x_t)
+                else:
+                    noise = 0
+                x_t = mean + torch.sqrt(var) * noise
+                assert torch.isnan(x_t).int().sum() == 0, "nan in tensor."
+        x_0 = x_t
+        return x_0
+
+
+if __name__ == "__main__":
+    from schedule import Schedule
+
+    test = GaussianDiffusion(None, 100, Schedule("linear", 100))
+    print(test.gammas)

schedule/schedule.py

@@ -0,0 +1,53 @@
+import numpy as np
+import torch
+
+
+class Schedule:
+    def __init__(self, schedule, timesteps):
+        self.timesteps = timesteps
+        self.schedule = schedule
+
+    def cosine_beta_schedule(self, s=0.001):
+        timesteps = self.timesteps
+        steps = timesteps + 1
+        x = torch.linspace(0, timesteps, steps)
+        alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * np.pi * 0.5) ** 2
+        alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
+        betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
+        return torch.clip(betas, 0.0001, 0.9999)
+
+    def linear_beta_schedule(self):
+        timesteps = self.timesteps
+        scale = 1000 / timesteps
+        beta_start = 1e-6 * scale
+        beta_end = 0.02 * scale
+        return torch.linspace(beta_start, beta_end, timesteps)
+
+    def quadratic_beta_schedule(self):
+        timesteps = self.timesteps
+        scale = 1000 / timesteps
+        beta_start = 1e-6 * scale
+        beta_end = 0.02 * scale
+        return torch.linspace(beta_start ** 0.5, beta_end ** 0.5, timesteps) ** 2
+
+    def sigmoid_beta_schedule(self):
+        timesteps = self.timesteps
+        scale = 1000 / timesteps
+        beta_start = 1e-6 * scale
+        beta_end = 0.02 * scale
+        betas = torch.linspace(-6, 6, timesteps)
+        return torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
+
+    def get_betas(self):
+        if self.schedule == "linear":
+            return self.linear_beta_schedule()
+        elif self.schedule == 'cosine':
+            return self.cosine_beta_schedule()
+        else:
+            raise NotImplementedError
+
+
+if __name__ == "__main__":
+    schedule = Schedule(schedule="linear", timesteps=100)
+    print(schedule.get_betas().shape)
+    print(schedule.get_betas())

src/config.py

@@ -0,0 +1,29 @@
+import yaml
+import os
+
+
+class Config(dict):
+    def __init__(self, config_path):
+        with open(config_path, 'r') as f:
+            self._yaml = f.read()
+            self._dict = yaml.safe_load(self._yaml)
+            self._dict['PATH'] = os.path.dirname(config_path)
+
+    def __getattr__(self, name):
+        if self._dict.get(name) is not None:
+            return self._dict[name]
+        return None
+
+    def print(self):
+        print('Model configurations:')
+        print('---------------------------------')
+        print(self._yaml)
+        print('')
+        print('---------------------------------')
+        print('')
+
+
+def load_config(path):
+    config_path = path
+    config = Config(config_path)
+    return config

src/sobel.py

@@ -0,0 +1,43 @@
+from torch import nn
+import torch
+
+
+class Sobel(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.filter = nn.Conv2d(in_channels=1, out_channels=2, kernel_size=3, stride=1, padding=1, bias=False)
+
+        Gx = torch.tensor([[1.0, 0.0, -1.0], [2.0, 0.0, -2.0], [1.0, 0.0, -1.0]])
+        Gy = torch.tensor([[1.0, 2.0, 1.0], [0.0, 0.0, 0.0], [-1.0, -2.0, -1.0]])
+        G = torch.cat([Gx.unsqueeze(0), Gy.unsqueeze(0)], 0)
+        G = G.unsqueeze(1)
+        self.filter.weight = nn.Parameter(G, requires_grad=False)
+
+    def forward(self, img):
+        if img.shape[1] == 3:
+            img = torch.mean(img, dim=1, keepdim=True)
+        x = self.filter(img)
+        x = torch.mul(x, x)
+        x = torch.sum(x, dim=1, keepdim=True)
+        x = torch.sqrt(x)
+        return x
+
+
+class Laplacian(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.filter = nn.Conv2d(in_channels=3, out_channels=3, kernel_size=3, stride=1, padding=1, bias=False, groups=3)
+        G = torch.tensor([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]]).float()
+        G = G.unsqueeze(0).unsqueeze(0)
+        G = torch.cat([G, G, G], 0)
+        self.filter.weight = nn.Parameter(G, requires_grad=False)
+
+    def forward(self, img):
+        x = self.filter(img)
+        return x
+
+if __name__ == "__main__":
+    laplacian = Laplacian()
+    img = torch.randn(1, 3, 256, 256)
+    y = laplacian(img)
+    print(y.shape)

src/train.py

@@ -0,0 +1,13 @@
+from src.trainer import Trainer
+
+
+def train(config):
+    trainer = Trainer(config)
+    trainer.train()
+    print('training complete')
+
+
+def test(config):
+    trainer = Trainer(config)
+    trainer.test()
+    print('testing complete')

src/trainer.py

@@ -0,0 +1,614 @@
+import copy
+import os
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from model.DocDiff import EMA, DocDiff
+from schedule.diffusionSample import GaussianDiffusion
+from schedule.dpm_solver_pytorch import DPM_Solver, NoiseScheduleVP, model_wrapper
+from schedule.schedule import Schedule
+from src.sobel import Laplacian
+from torch.utils.data import DataLoader
+from torchvision.transforms import Resize
+from torchvision.utils import save_image
+from tqdm import tqdm
+from utils.perceptual_loss import PerceptualLoss
+from utils.utils import get_A
+
+# from utils.RGBuvHistBlock import RGBuvHistBlock
+# from depth_anything.dpt import DPT_DINOv2
+
+
+class Trainer:
+    def __init__(self, config):
+        self.mode = config.MODE
+        self.schedule = Schedule(config.SCHEDULE, config.TIMESTEPS)
+        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        in_channels = config.CHANNEL_X + config.CHANNEL_Y
+        out_channels = config.CHANNEL_Y
+        self.out_channels = out_channels
+        self.network = DocDiff(
+            input_channels=in_channels,
+            output_channels=out_channels,
+            n_channels=config.MODEL_CHANNELS,
+            ch_mults=config.CHANNEL_MULT,
+            n_blocks=config.NUM_RESBLOCKS,
+        ).to(self.device)
+        self.diffusion = GaussianDiffusion(
+            self.network.denoiser, config.TIMESTEPS, self.schedule
+        ).to(self.device)
+        self.output_dir = config.OUTPUT_DIR
+        self.test_path = os.path.join(config.OUTPUT_DIR, config.TEST_PATH)
+        if not os.path.exists(self.test_path):
+            os.makedirs(self.test_path)
+        self.vis_path = os.path.join(config.OUTPUT_DIR, config.VIS_PATH)
+        if not os.path.exists(self.vis_path):
+            os.makedirs(self.vis_path)
+        self.train_path = os.path.join(config.OUTPUT_DIR, config.TRAIN_PATH)
+        if not os.path.exists(self.train_path):
+            os.makedirs(self.train_path)
+        self.weight_save_path = os.path.join(config.OUTPUT_DIR, config.WEIGHT_SAVE_PATH)
+        if not os.path.exists(self.weight_save_path):
+            os.makedirs(self.weight_save_path)
+
+        self.pretrained_path_beta_predictor = config.PRETRAINED_PATH_BETA_PREDICTOR
+        self.pretrained_path_denoiser = config.PRETRAINED_PATH_DENOISER
+        self.pretrained_path_depth_estimator = config.PRETRAINED_PATH_DEPTH_ESTIMATOR
+
+        self.continue_training = config.CONTINUE_TRAINING
+        self.continue_training_steps = 0
+
+        self.path_train_gt = config.PATH_GT
+        self.path_train_img = config.PATH_IMG
+        self.path_train_label_depth = config.PATH_GT_DEPTH
+        self.path_train_hist = config.PATH_IMG_HIST
+
+        self.path_test_img = config.PATH_TEST_IMG
+        self.path_test_gt = config.PATH_TEST_GT
+        self.path_test_label_depth = config.PATH_TEST_GT_DEPTH
+        self.path_test_hist = config.PATH_TEST_IMG_HIST
+
+        self.beta_loss = config.BETA_LOSS
+        self.pre_ori = config.PRE_ORI
+        self.high_low_freq = config.HIGH_LOW_FREQ
+        self.image_size = config.IMAGE_SIZE
+        self.native_resolution = config.NATIVE_RESOLUTION
+        self.iteration_max = config.ITERATION_MAX
+        self.LR = config.LR
+        self.cross_entropy = nn.BCELoss()
+        self.num_timesteps = config.TIMESTEPS
+        self.ema_every = config.EMA_EVERY
+        self.start_ema = config.START_EMA
+        self.save_model_every = config.SAVE_MODEL_EVERY
+        self.EMA_or_not = config.EMA
+        self.DPM_SOLVER = config.DPM_SOLVER
+        self.DPM_STEP = config.DPM_STEP
+
+        if self.mode == 1 and self.continue_training == "True":
+            print("Continue Training")
+            self.network.beta_predictor.load_state_dict(
+                torch.load(self.pretrained_path_beta_predictor)
+            )
+            self.network.denoiser.load_state_dict(
+                torch.load(self.pretrained_path_denoiser)
+            )
+            self.continue_training_steps = config.CONTINUE_TRAINING_STEPS
+
+        # if self.mode == 0:
+        #     self.depth_estimator = DPT_DINOv2(
+        #         encoder="vits",
+        #         features=64,
+        #         out_channels=[48, 96, 192, 384],
+        #         localhub=True)
+        #     self.depth_estimator.load_state_dict(
+        #     torch.load(
+        #         self.pretrained_path_depth_estimator,
+        #         map_location="cpu",
+        #     ),
+        #     strict=True,
+        #     )
+        #     self.hist_estimator = RGBuvHistBlock(
+        #         insz=config.IMAGE_SIZE[0],
+        #         h=config.IMAGE_SIZE[1],
+        #         resizing='sampling',
+        #         method='inverse-quadratic',
+        #         sigma=0.02,
+        #         device=self.device)
+
+        from data.data import UIEData
+
+        if self.mode == 1:
+            dataset_train = UIEData(
+                self.path_train_img,
+                self.path_train_gt,
+                self.path_train_label_depth,
+                self.path_train_hist,
+                config.IMAGE_SIZE,
+                self.mode,
+            )
+            self.dataloader_train = DataLoader(
+                dataset_train,
+                batch_size=config.BATCH_SIZE,
+                shuffle=True,
+                drop_last=False,
+                num_workers=config.NUM_WORKERS,
+            )
+            dataset_eval = UIEData(
+                self.path_test_img,
+                self.path_test_gt,
+                self.path_test_label_depth,
+                self.path_test_hist,
+                config.IMAGE_SIZE,
+                mode=0,
+            )
+            self.dataloader_eval = DataLoader(
+                dataset_eval,
+                batch_size=config.BATCH_SIZE_VAL,
+                shuffle=False,
+                drop_last=False,
+                num_workers=config.NUM_WORKERS,
+            )
+        else:
+            dataset_test = UIEData(
+                self.path_test_img,
+                self.path_test_gt,
+                self.path_test_label_depth,
+                self.path_test_hist,
+                config.IMAGE_SIZE,
+                self.mode,
+            )
+            self.dataloader_test = DataLoader(
+                dataset_test,
+                batch_size=config.BATCH_SIZE_VAL,
+                shuffle=False,
+                drop_last=False,
+                num_workers=config.NUM_WORKERS,
+            )
+        if self.mode == 1 and config.EMA == "True":
+            self.EMA = EMA(0.9999)
+            self.ema_model = copy.deepcopy(self.network).to(self.device)
+        if config.LOSS == "L1":
+            self.loss = nn.L1Loss()
+        elif config.LOSS == "L2":
+            self.loss = nn.MSELoss()
+        else:
+            print("Loss not implemented, setting the loss to L2 (default one)")
+            self.loss = nn.MSELoss()
+        if self.high_low_freq == "True":
+            self.high_filter = Laplacian().to(self.device)
+        self.perceptual_loss = PerceptualLoss()
+
+    def test(self):
+        def crop_concat(img, size=128):
+            shape = img.shape
+            correct_shape = (
+                size * (shape[2] // size + 1),
+                size * (shape[3] // size + 1),
+            )
+            one = torch.ones((shape[0], shape[1], correct_shape[0], correct_shape[1]))
+            one[:, :, : shape[2], : shape[3]] = img
+            # crop
+            for i in range(shape[2] // size + 1):
+                for j in range(shape[3] // size + 1):
+                    if i == 0 and j == 0:
+                        crop = one[
+                            :, :, i * size : (i + 1) * size, j * size : (j + 1) * size
+                        ]
+                    else:
+                        crop = torch.cat(
+                            (
+                                crop,
+                                one[
+                                    :,
+                                    :,
+                                    i * size : (i + 1) * size,
+                                    j * size : (j + 1) * size,
+                                ],
+                            ),
+                            dim=0,
+                        )
+            return crop
+
+        def crop_concat_back(img, prediction, size=128):
+            shape = img.shape
+            for i in range(shape[2] // size + 1):
+                for j in range(shape[3] // size + 1):
+                    if j == 0:
+                        crop = prediction[
+                            (i * (shape[3] // size + 1) + j) * shape[0] : (
+                                i * (shape[3] // size + 1) + j + 1
+                            )
+                            * shape[0],
+                            :,
+                            :,
+                            :,
+                        ]
+                    else:
+                        crop = torch.cat(
+                            (
+                                crop,
+                                prediction[
+                                    (i * (shape[3] // size + 1) + j) * shape[0] : (
+                                        i * (shape[3] // size + 1) + j + 1
+                                    )
+                                    * shape[0],
+                                    :,
+                                    :,
+                                    :,
+                                ],
+                            ),
+                            dim=3,
+                        )
+                if i == 0:
+                    crop_concat = crop
+                else:
+                    crop_concat = torch.cat((crop_concat, crop), dim=2)
+            return crop_concat[:, :, : shape[2], : shape[3]]
+
+        def min_max(array):
+            return (array - array.min()) / (array.max() - array.min())
+
+        with torch.no_grad():
+            self.network.beta_predictor.load_state_dict(
+                torch.load(self.pretrained_path_beta_predictor)
+            )
+            self.network.denoiser.load_state_dict(
+                torch.load(self.pretrained_path_denoiser)
+            )
+            print("Test Model loaded")
+            self.network.eval()
+            tq = tqdm(self.dataloader_test)
+            sampler = self.diffusion
+            iteration = 0
+            num_iter = 10
+            for img, gt, label_depth, hist, name, in_size in tq:
+                tq.set_description(
+                    f"Iteration {iteration} / {len(self.dataloader_test.dataset)}"
+                )
+                iteration += 1
+                pred_beta = self.network.beta_predictor(img.to(self.device))
+                T_direct = torch.clamp((torch.exp(-pred_beta * label_depth)), 0, 1)
+                T_scatter = torch.clamp((1 - torch.exp(-pred_beta * label_depth)), 0, 1)
+                atm_light = [get_A(item) for item in img.to(self.device)]
+                atm_light = torch.stack(atm_light).to(self.device)
+                J = torch.clamp(
+                    ((img.to(self.device) - T_scatter * atm_light) / T_direct), 0, 1
+                )
+
+                noisyImage = torch.randn_like(img.to(self.device))
+
+                if self.DPM_SOLVER == "True":
+                    sampledImgs = dpm_solver(
+                        self.schedule.get_betas(),
+                        self.network,
+                        torch.cat((noisyImage, img.to(self.device)), dim=1),
+                        self.DPM_STEP,
+                    )
+                else:
+                    sampledImgs = sampler(
+                        noisyImage.cuda(),
+                        img.to(self.device),
+                        J.to(self.device),
+                        hist.to(self.device),
+                        self.pre_ori,
+                    )
+                img_save = torch.cat(
+                    [img, gt, J.cpu(), hist.cpu(), sampledImgs.cpu()], dim=3
+                )
+
+                if not os.path.exists(os.path.join(self.test_path, f"{num_iter}")):
+                    os.makedirs(os.path.join(self.test_path, f"{num_iter}"))
+                save_image(
+                    img_save,
+                    os.path.join(self.test_path, f"{num_iter}", f"{name[0]}"),
+                    nrow=4,
+                )
+                if not os.path.exists(os.path.join(self.vis_path, f"{num_iter}")):
+                    os.makedirs(os.path.join(self.vis_path, f"{num_iter}"))
+                save_image(
+                    Resize(in_size)(sampledImgs).cpu(),
+                    os.path.join(self.vis_path, f"{num_iter}", f"{name[0]}"),
+                )
+
+    def evaluation(self, num_iter, beta_predictor_weight, denoiser_weight):
+        def crop_concat(img, size=128):
+            shape = img.shape
+            correct_shape = (
+                size * (shape[2] // size + 1),
+                size * (shape[3] // size + 1),
+            )
+            one = torch.ones((shape[0], shape[1], correct_shape[0], correct_shape[1]))
+            one[:, :, : shape[2], : shape[3]] = img
+            # crop
+            for i in range(shape[2] // size + 1):
+                for j in range(shape[3] // size + 1):
+                    if i == 0 and j == 0:
+                        crop = one[
+                            :, :, i * size : (i + 1) * size, j * size : (j + 1) * size
+                        ]
+                    else:
+                        crop = torch.cat(
+                            (
+                                crop,
+                                one[
+                                    :,
+                                    :,
+                                    i * size : (i + 1) * size,
+                                    j * size : (j + 1) * size,
+                                ],
+                            ),
+                            dim=0,
+                        )
+            return crop
+
+        def crop_concat_back(img, prediction, size=128):
+            shape = img.shape
+            for i in range(shape[2] // size + 1):
+                for j in range(shape[3] // size + 1):
+                    if j == 0:
+                        crop = prediction[
+                            (i * (shape[3] // size + 1) + j) * shape[0] : (
+                                i * (shape[3] // size + 1) + j + 1
+                            )
+                            * shape[0],
+                            :,
+                            :,
+                            :,
+                        ]
+                    else:
+                        crop = torch.cat(
+                            (
+                                crop,
+                                prediction[
+                                    (i * (shape[3] // size + 1) + j) * shape[0] : (
+                                        i * (shape[3] // size + 1) + j + 1
+                                    )
+                                    * shape[0],
+                                    :,
+                                    :,
+                                    :,
+                                ],
+                            ),
+                            dim=3,
+                        )
+                if i == 0:
+                    crop_concat = crop
+                else:
+                    crop_concat = torch.cat((crop_concat, crop), dim=2)
+            return crop_concat[:, :, : shape[2], : shape[3]]
+
+        def min_max(array):
+            return (array - array.min()) / (array.max() - array.min())
+
+        with torch.no_grad():
+            self.network.beta_predictor.load_state_dict(
+                torch.load(beta_predictor_weight)
+            )
+            self.network.denoiser.load_state_dict(torch.load(denoiser_weight))
+
+            print("Eval Model loaded")
+            self.network.eval()
+            tq = tqdm(self.dataloader_eval)
+            sampler = self.diffusion
+            iteration = 0
+            for img, gt, label_depth, hist, name, in_size in tq:
+                tq.set_description(
+                    f"Iteration {iteration} / {len(self.dataloader_eval.dataset)}"
+                )
+                iteration += 1
+                if self.native_resolution == "True":
+                    temp = img
+                    img = crop_concat(img)
+                pred_beta = self.network.beta_predictor(img.to(self.device))
+                depth = label_depth.to(self.device)
+                T_direct = torch.clamp((torch.exp(-pred_beta * depth)), 0, 1)
+                T_scatter = torch.clamp((1 - torch.exp(-pred_beta * depth)), 0, 1)
+                atm_light = [get_A(item) for item in img.to(self.device)]
+                atm_light = torch.stack(atm_light).to(self.device)
+                J = torch.clamp(
+                    ((img.to(self.device) - T_scatter * atm_light) / T_direct), 0, 1
+                )
+                noisyImage = torch.randn_like(img.to(self.device))
+
+                if self.DPM_SOLVER == "True":
+                    sampledImgs = dpm_solver(
+                        self.schedule.get_betas(),
+                        self.network,
+                        torch.cat((noisyImage, img.to(self.device)), dim=1),
+                        self.DPM_STEP,
+                    )
+                else:
+                    sampledImgs = sampler(
+                        noisyImage.cuda(),
+                        img.to(self.device),
+                        J.to(self.device),
+                        hist.to(self.device),
+                        self.pre_ori,
+                    )
+                img_save = torch.cat(
+                    [
+                        img,
+                        gt,
+                        J.cpu(),
+                        hist.cpu(),
+                        sampledImgs.cpu(),
+                        T_direct.cpu(),
+                        T_scatter.cpu(),
+                        label_depth.cpu(),
+                    ],
+                    dim=3,
+                )
+
+                if not os.path.exists(os.path.join(self.test_path, f"{num_iter}")):
+                    os.makedirs(os.path.join(self.test_path, f"{num_iter}"))
+                save_image(
+                    img_save,
+                    os.path.join(self.test_path, f"{num_iter}", f"{name[0]}"),
+                    nrow=4,
+                )
+                if not os.path.exists(os.path.join(self.vis_path, f"{num_iter}")):
+                    os.makedirs(os.path.join(self.vis_path, f"{num_iter}"))
+                save_image(
+                    Resize(in_size)(sampledImgs).cpu(),
+                    os.path.join(self.vis_path, f"{num_iter}", f"{name[0]}"),
+                )
+
+    def train(self):
+        optimizer = optim.AdamW(
+            self.network.parameters(), lr=self.LR, weight_decay=1e-4
+        )
+        iteration = self.continue_training_steps
+        print("Starting Training", f"Step is {self.num_timesteps}")
+        total_params = sum(
+            p.numel() for p in self.network.parameters() if p.requires_grad
+        )
+        print(f"Trainable parameters: {total_params}")
+        while iteration < self.iteration_max:
+            tq = tqdm(self.dataloader_train)
+
+            for img, gt, label_depth, hist, _, _ in tq:
+                tq.set_description(f"Iteration {iteration} / {self.iteration_max}")
+                self.network.train()
+                optimizer.zero_grad()
+
+                t = (
+                    torch.randint(0, self.num_timesteps, (img.shape[0],))
+                    .long()
+                    .to(self.device)
+                )
+                J, noise_ref, denoised_J, T_direct, T_scatter = self.network(
+                    gt.to(self.device),
+                    img.to(self.device),
+                    hist.to(self.device),
+                    label_depth.to(self.device),
+                    t,
+                    self.diffusion,
+                )
+                if self.pre_ori == "True":
+                    ddpm_loss = self.loss(denoised_J, gt.to(self.device))
+                    perceptual_loss = self.perceptual_loss(
+                        denoised_J, gt.to(self.device)
+                    )
+                else:
+                    ddpm_loss = self.loss(denoised_J, noise_ref.to(self.device))
+                    perceptual_loss = self.perceptual_loss(
+                        denoised_J, noise_ref.to(self.device)
+                    )
+
+                loss = ddpm_loss + perceptual_loss
+                loss.backward()
+                optimizer.step()
+
+                tq.set_postfix(
+                    loss=loss.item(),
+                    ddpm_loss=ddpm_loss.item(),
+                    perceptual_loss=perceptual_loss.item(),
+                )
+
+                if iteration % 1000 == 0:
+                    img_save = torch.cat(
+                        [
+                            img,
+                            gt,
+                            J.cpu(),
+                            denoised_J.cpu(),
+                            hist.cpu(),
+                            T_direct.cpu(),
+                            T_scatter.cpu(),
+                            label_depth.cpu(),
+                        ],
+                        dim=3,
+                    )
+                    save_image(
+                        img_save,
+                        os.path.join(self.train_path, f"{iteration}.png"),
+                        nrow=4,
+                    )
+                iteration += 1
+                if self.EMA_or_not == "True":
+                    if iteration % self.ema_every == 0 and iteration > self.start_ema:
+                        print("EMA update")
+                        self.EMA.update_model_average(self.ema_model, self.network)
+
+                if iteration % self.save_model_every == 0:
+                    print("Saving models")
+                    if not os.path.exists(self.weight_save_path):
+                        os.makedirs(self.weight_save_path)
+                    torch.save(
+                        self.network.beta_predictor.state_dict(),
+                        os.path.join(
+                            self.weight_save_path,
+                            f"model_beta_predictor_{iteration}.pth",
+                        ),
+                    )
+                    torch.save(
+                        self.network.denoiser.state_dict(),
+                        os.path.join(
+                            self.weight_save_path, f"model_denoiser_{iteration}.pth"
+                        ),
+                    )
+
+                    self.evaluation(
+                        iteration,
+                        os.path.join(
+                            self.weight_save_path,
+                            f"model_beta_predictor_{iteration}.pth",
+                        ),
+                        os.path.join(
+                            self.weight_save_path, f"model_denoiser_{iteration}.pth"
+                        ),
+                    )
+
+
+def dpm_solver(betas, model, x_T, steps, model_kwargs):
+    # You need to firstly define your model and the extra inputs of your model,
+    # And initialize an `x_T` from the standard normal distribution.
+    # `model` has the format: model(x_t, t_input, **model_kwargs).
+    # If your model has no extra inputs, just let model_kwargs = {}.
+
+    # If you use discrete-time DPMs, you need to further define the
+    # beta arrays for the noise schedule.
+
+    # model = ....
+    # model_kwargs = {...}
+    # x_T = ...
+    # betas = ....
+
+    # 1. Define the noise schedule.
+    noise_schedule = NoiseScheduleVP(schedule="discrete", betas=betas)
+
+    # 2. Convert your discrete-time `model` to the continuous-time
+    # noise prediction model. Here is an example for a diffusion model
+    # `model` with the noise prediction type ("noise") .
+    model_fn = model_wrapper(
+        model,
+        noise_schedule,
+        model_type="noise",  # or "x_start" or "v" or "score"
+        model_kwargs=model_kwargs,
+    )
+
+    # 3. Define dpm-solver and sample by singlestep DPM-Solver.
+    # (We recommend singlestep DPM-Solver for unconditional sampling)
+    # You can adjust the `steps` to balance the computation
+    # costs and the sample quality.
+    dpm_solver = DPM_Solver(
+        model_fn,
+        noise_schedule,
+        algorithm_type="dpmsolver++",
+        correcting_x0_fn="dynamic_thresholding",
+    )
+    # Can also try
+    # dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+
+    # You can use steps = 10, 12, 15, 20, 25, 50, 100.
+    # Empirically, we find that steps in [10, 20] can generate quite good samples.
+    # And steps = 20 can almost converge.
+    x_sample = dpm_solver.sample(
+        x_T,
+        steps=steps,
+        order=1,
+        skip_type="time_uniform",
+        method="singlestep",
+    )
+    return x_sample

torchhub/README.md

@@ -0,0 +1,3 @@
+# Local PyTorch Hub
+
+This directory is for loading the DINOv2 encoder locally in case of no Internet connection.

torchhub/facebookresearch_dinov2_main/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <opensource-conduct@meta.com>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

torchhub/facebookresearch_dinov2_main/CONTRIBUTING.md

@@ -0,0 +1,31 @@
+# Contributing to DINOv2
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Meta's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Meta has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to DINOv2, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

torchhub/facebookresearch_dinov2_main/LICENSE

@@ -0,0 +1,400 @@
+
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public: 
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

torchhub/facebookresearch_dinov2_main/MODEL_CARD.md

@@ -0,0 +1,201 @@
+# Model Card for DINOv2-S/B/L/g
+
+These are Vision Transformer models trained following the method described in the paper:
+"DINOv2: Learning Robust Visual Features without Supervision"
+
+We provide 4 models: 1 ViT-g trained from scratch, and 3 ViT-S/B/L models distilled from the ViT-g.
+
+## Model Details
+The model takes an image as input and returns a class token and patch tokens.
+
+The embedding dimension is: 
+- 384 for ViT-S.
+- 768 for ViT-B.
+- 1024 for ViT-L.
+- 1536 for ViT-g.
+
+The models follow a Transformer architecture, with a patch size of 14.
+
+For a 224x224 image, this results in 1 class token + 256 patch tokens.
+
+The models can accept larger images provided the image shapes are multiples of the patch size (14). 
+If this condition is not verified, the model will crop to the closest smaller multiple of the patch size.
+
+### Model Description
+
+- **Developed by:** Meta AI
+- **Model type:** Vision Transformer
+- **License:** CC-BY-NC
+
+- **Repository:** https://github.com/facebookresearch/dinov2
+- **Paper:** https://arxiv.org/abs/2304.07193
+- **Demo:** https://dinov2.metademolab.com/
+
+## Uses
+
+The models are vision backbones providing multi-purpose features for downstream tasks.
+
+### Direct Use
+
+The models can be used without fine-tuning, with downstream classifiers as simple as linear layers, to obtain competitive results:
+- on depth estimation, semantic segmentation, using linear layers.
+- on image classification, using k-NN classifiers on the class token.
+- on image classification, with logistic regression classifiers applied on the class token.
+- on image classification, with a linear layer applied on the class token and the average of the patch tokens.
+- on image retrieval using nearest neighbors.
+
+### Downstream Use
+
+It is technically possible to perform fine-tuning on the models, for small gains (we measured +2% on ImageNet-1k classification). 
+We recommend keeping this as a very last step and only when necessary, as the features already provide good performance out-of-the-box.
+
+## Bias, Risks, and Limitations
+
+Despite improvements thanks to the training method not using annotations, we still observe significant biases in our models toward rich households from Western countries.
+
+### Recommendations
+
+We expect fine-tuning will increase the biases in the features produced by the model as they will be tuned to the fine-tuning labels.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+```python
+import torch
+dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
+dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
+dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
+dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')
+```
+
+## Training Details
+
+### Training Data
+
+- **Training data:** LVD-142M (see paper)
+- **Training regime:** fp16 using PyTorch-FSDP mixed-precision.
+
+### Training Procedure 
+
+- **Training objective:**
+  - DINO self-distillation loss with multi-crop
+  - iBOT masked-image modeling loss
+  - KoLeo regularization on [CLS] tokens
+- **Architectures:**
+  - ViT-S (21M params): Patch size 14, embedding dimension 384, 6 heads, MLP FFN
+  - ViT-B (86M params): Patch size 14, embedding dimension 768, 12 heads, MLP FFN
+  - ViT-L (0.3B params): Patch size 14, embedding dimension 1024, 16 heads, MLP FFN
+  - ViT-g (1.1B params): Patch size 14, embedding dimension 1536, 24 heads, SwiGLU FFN
+- **Distillation:**
+  - Distillation follows the standard DINOv2 pretraining procedure, except the teacher is a pretrained ViT-g, frozen.
+
+## Evaluation
+
+We refer users to the associated paper for the evaluation protocols.
+
+<table>
+  <tr>
+    <th>model</th>
+    <th colspan="3">ImageNet-1k</th>
+    <th>NYU-Depth v2</th>
+    <th>SUN-RGBD</th>
+    <th>ADE20k</th>
+    <th>iNaturalist 2018</th>
+    <th>Oxford-H</th>
+  </tr>
+  <tr>
+    <th rowspan="2">task</th>
+    <th>classif. (acc)</th>
+    <th>classif. (acc)</th>
+    <th>classif. V2 (acc)</th>
+    <th>depth (RMSE)</th>
+    <th>depth (RMSE)</th>
+    <th>segm. (mAP)</th>
+    <th>classif. (acc)</th>
+    <th>retrieval (mAP)</th>
+  </tr>
+  <tr>
+    <!-- <th>^</th> -->
+    <th>k-NN</th>
+    <th>linear</th>
+    <th>linear</th>
+    <th>linear<br />4 layers</th>
+    <th>NYU-D transfer</th>
+    <th>multiscale</th>
+    <th>linear</th>
+    <th>nearest neighbor</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14</td>
+    <td align="right">79.0%</td>
+    <td align="right">81.1%</td>
+    <td align="right">70.8%</td> 
+    <td align="right">0.417</td> 
+    <td align="right">0.431</td> 
+    <td align="right">47.2</td> 
+    <td align="right">69.5%</td> 
+    <td align="right">43.2</td> 
+  </tr>
+  <tr>
+    <td>ViT-B/14</td>
+    <td align="right">82.1%</td>
+    <td align="right">84.5%</td>
+    <td align="right">74.9%</td>
+    <td align="right">0.362</td> 
+    <td align="right">0.400</td> 
+    <td align="right">51.3</td> 
+    <td align="right">76.3%</td> 
+    <td align="right">49.5</td> 
+  </tr>
+  <tr>
+    <td>ViT-L/14</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.3%</td>
+    <td align="right">77.6%</td>
+    <td align="right">0.333</td> 
+    <td align="right">0.396</td> 
+    <td align="right">53.1</td> 
+    <td align="right">79.8%</td> 
+    <td align="right">54.0</td> 
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.5%</td>
+    <td align="right">78.4%</td>
+    <td align="right">0.298</td> 
+    <td align="right">0.362</td> 
+    <td align="right">53.0</td> 
+    <td align="right">81.6%</td> 
+    <td align="right">52.3</td> 
+  </tr>
+</table>
+
+## Environmental Impact
+
+- **Hardware Type:** Nvidia A100
+- **Hours used:** 22,000 for ViT-g, 4,500 for ViT-S distillation, 5,300 for ViT-B distillation, 8,000 for ViT-L distillation
+- **Cloud Provider:** Private infra
+- **Compute Region:** USA
+- **Carbon Emitted:** 7t CO2eq
+
+#### Hardware
+
+Nvidia A100 GPUs
+
+#### Software
+
+PyTorch 2.0,
+xFormers 0.0.18
+
+**BibTeX**
+
+```
+@misc{oquab2023dinov2,
+  title={DINOv2: Learning Robust Visual Features without Supervision},
+  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
+  journal={arXiv:2304.07193},
+  year={2023}
+}
+```

torchhub/facebookresearch_dinov2_main/README.md

@@ -0,0 +1,277 @@
+# DINOv2: Learning Robust Visual Features without Supervision
+
+**[Meta AI Research, FAIR](https://ai.facebook.com/research/)**
+
+Maxime Oquab,
+Timothée Darcet,
+Théo Moutakanni,
+Huy V. Vo,
+Marc Szafraniec,
+Vasil Khalidov,
+Patrick Labatut,
+Armand Joulin,
+Piotr Bojanowski
+
+[[`Paper`](https://arxiv.org/abs/2304.07193)] [[`Blog`](https://ai.facebook.com/blog/dino-v2-computer-vision-self-supervised-learning/)] [[`Demo`](https://dinov2.metademolab.com)] [[`BibTeX`](#citing-dinov2)]
+
+PyTorch implementation and pretrained models for DINOv2. For details, see the paper: **[DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)**.
+
+DINOv2 models produce high-performance visual features that can be directly employed with classifiers as simple as linear layers on a variety of computer vision tasks; these visual features are robust and perform well across domains without any requirement for fine-tuning. The models were pretrained on a dataset of 142 M images without using any labels or annotations.
+
+https://github.com/facebookresearch/dinov2/assets/60359573/f168823e-7922-415a-b429-578badf5c356
+
+<div align="center">
+  Visualization of the three first principal components of the patch features of all frames, mapped to RGB values.
+</div>
+
+## Pretrained models
+
+<table style="margin: auto">
+  <tr>
+    <th>model</th>
+    <th># of<br />params</th>
+    <th>ImageNet<br />k-NN</th>
+    <th>ImageNet<br />linear</th>
+    <th>download</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14 distilled</td>
+    <td align="right">21 M</td>
+    <td align="right">79.0%</td>
+    <td align="right">81.1%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-B/14 distilled</td>
+    <td align="right">86 M</td>
+    <td align="right">82.1%</td>
+    <td align="right">84.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-L/14 distilled</td>
+    <td align="right">300 M</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.3%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_pretrain.pth">backbone only</a></td>
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">1,100 M</td>
+    <td align="right">83.5%</td>
+    <td align="right">86.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_pretrain.pth">backbone only</a></td>
+  </tr>
+</table>
+
+### Pretrained models via PyTorch Hub
+
+Please follow the instructions [here](https://pytorch.org/get-started/locally/) to install PyTorch (the only required dependency for loading the model). Installing PyTorch with CUDA support is strongly recommended.
+
+A corresponding [model card](MODEL_CARD.md) is included in the repository.
+
+```python
+import torch
+
+dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
+dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
+dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
+dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')
+```
+
+## Installation
+
+The training and evaluation code requires PyTorch 2.0 and [xFormers](https://github.com/facebookresearch/xformers) 0.0.18 as well as a number of other 3rd party packages. Note that the code has only been tested with the specified versions and also expects a Linux environment. To setup all the required dependencies for training and evaluation, please follow the instructions below:
+
+*[conda](https://docs.conda.io/projects/conda/en/latest/user-guide/getting-started.html)* **(Recommended)** - Clone the repository and then create and activate a `dinov2` conda environment using the provided environment definition:
+
+```shell
+conda env create -f conda.yaml
+conda activate dinov2
+```
+
+*[pip](https://pip.pypa.io/en/stable/getting-started/)* - Clone the repository and then use the provided `requirements.txt` to install the dependencies:
+
+```shell
+pip install -r requirements.txt
+```
+
+## Data preparation
+
+### ImageNet-1k
+
+The root directory of the dataset should hold the following contents:
+
+- `<ROOT>/test/ILSVRC2012_test_00000001.JPEG`
+- `<ROOT>/test/[..]`
+- `<ROOT>/test/ILSVRC2012_test_00100000.JPEG`
+- `<ROOT>/train/n01440764/n01440764_10026.JPEG`
+- `<ROOT>/train/[...]`
+- `<ROOT>/train/n15075141/n15075141_9993.JPEG`
+- `<ROOT>/val/n01440764/ILSVRC2012_val_00000293.JPEG`
+- `<ROOT>/val/[...]`
+- `<ROOT>/val/n15075141/ILSVRC2012_val_00049174.JPEG`
+- `<ROOT>/labels.txt`
+
+The provided dataset implementation expects a few additional metadata files to be present under the extra directory:
+
+- `<EXTRA>/class-ids-TRAIN.npy`
+- `<EXTRA>/class-ids-VAL.npy`
+- `<EXTRA>/class-names-TRAIN.npy`
+- `<EXTRA>/class-names-VAL.npy`
+- `<EXTRA>/entries-TEST.npy`
+- `<EXTRA>/entries-TRAIN.npy`
+- `<EXTRA>/entries-VAL.npy`
+
+These metadata files can be generated (once) with the following lines of Python code:
+
+```python
+from dinov2.data.datasets import ImageNet
+
+for split in ImageNet.Split:
+    dataset = ImageNet(split=split, root="<ROOT>", extra="<EXTRA>")
+    dataset.dump_extra()
+```
+
+Note that the root and extra directories do not have to be distinct directories.
+
+### ImageNet-22k
+
+Please adapt the [dataset class](dinov2/data/datasets/image_net_22k.py) to match your local setup.
+
+<br />
+
+:warning: To execute the commands provided in the next sections for training and evaluation, the `dinov2` package should be included in the Python module search path, i.e. simply prefix the command to run with `PYTHONPATH=.`.
+
+## Training
+
+### Fast setup: training DINOv2 ViT-L/16 on ImageNet-1k
+
+Run DINOv2 training on 4 A100-80GB nodes (32 GPUs) in a SLURM cluster environment with submitit:
+
+```shell
+python dinov2/run/train/train.py \
+    --nodes 4 \
+    --config-file dinov2/configs/train/vitl16_short.yaml \
+    --output-dir <PATH/TO/OUTPUT/DIR> \
+    train.dataset_path=ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+Training time is approximately 1 day and the resulting checkpoint should reach 81.6% on k-NN eval and 82.9% on linear eval.
+
+The training code saves the weights of the teacher in the `eval` folder every 12500 iterations for evaluation.
+
+### Long setup: training DINOv2 ViT-L/14 on ImageNet-22k
+
+Run DINOv2 training on 12 A100-80GB nodes (96 GPUs) in a SLURM cluster environment with submitit:
+
+```shell
+python dinov2/run/train/train.py \
+    --nodes 12 \
+    --config-file dinov2/configs/train/vitl14.yaml \
+    --output-dir <PATH/TO/OUTPUT/DIR> \
+    train.dataset_path=ImageNet22k:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+Training time is approximately 3.3 days and the resulting checkpoint should reach 82.0% on k-NN eval and 84.5% on linear eval.
+
+The training code saves the weights of the teacher in the `eval` folder every 12500 iterations for evaluation.
+
+
+## Evaluation
+
+The training code regularly saves the teacher weights. In order to evaluate the model, run the following evaluation on a single node:
+
+### k-NN classification on ImageNet-1k
+
+```shell
+python dinov2/run/eval/knn.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/knn \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+### Logistic regression classification on ImageNet-1k
+
+```shell
+python dinov2/run/eval/log_regression.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/logreg \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+### Linear classification with data augmentation on ImageNet-1k
+
+```shell
+python dinov2/run/eval/linear.py \
+    --config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
+    --pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
+    --output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/linear \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+We release the weights from evaluating the different models:
+
+<table style="margin: auto">
+  <tr>
+    <th>model</th>
+    <th>ImageNet<br />top-1</th>
+    <th>linear evaluation</th>
+  </tr>
+  <tr>
+    <td>ViT-S/14 distilled</td>
+    <td align="right">81.1%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-B/14 distilled</td>
+    <td align="right">84.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-L/14 distilled</td>
+    <td align="right">86.3%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_linear_head.pth">linear head weights</a></td>
+  </tr>
+  <tr>
+    <td>ViT-g/14</td>
+    <td align="right">86.5%</td>
+    <td><a href="https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_linear_head.pth">linear head weights</a></td>
+  </tr>
+</table>
+
+The performance of the provided pretrained model weights can be evaluated as follows on ImageNet-1k:
+
+```shell
+python dinov2/run/eval/linear.py \
+    --config-file dinov2/configs/eval/vitg14_pretrain.yaml \
+    --pretrained-weights https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_pretrain.pth \
+    --train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
+    --val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>
+```
+
+## License
+
+DINOv2 code and model weights are released under the CC-BY-NC 4.0 license. See [LICENSE](LICENSE) for additional details.
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+
+## Citing DINOv2
+
+If you find this repository useful, please consider giving a star :star: and citation :t-rex::
+
+```
+@misc{oquab2023dinov2,
+  title={DINOv2: Learning Robust Visual Features without Supervision},
+  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy V. and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
+  journal={arXiv:2304.07193},
+  year={2023}
+}
+```

torchhub/facebookresearch_dinov2_main/conda.yaml

@@ -0,0 +1,22 @@
+name: dinov2
+channels:
+  - defaults
+  - pytorch
+  - nvidia
+  - xformers
+  - conda-forge
+dependencies:
+  - python=3.9
+  - pytorch::pytorch=2.0.0
+  - pytorch::pytorch-cuda=11.7.0
+  - pytorch::torchvision=0.15.0
+  - omegaconf
+  - torchmetrics=0.10.3
+  - fvcore
+  - iopath
+  - xformers::xformers=0.0.18
+  - pip
+  - pip:
+    - git+https://github.com/facebookincubator/submitit
+    - --extra-index-url https://pypi.nvidia.com
+    - cuml-cu11

torchhub/facebookresearch_dinov2_main/dinov2/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+__version__ = "0.0.1"

torchhub/facebookresearch_dinov2_main/dinov2/configs/__init__.py

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pathlib
+
+from omegaconf import OmegaConf
+
+
+def load_config(config_name: str):
+    config_filename = config_name + ".yaml"
+    return OmegaConf.load(pathlib.Path(__file__).parent.resolve() / config_filename)
+
+
+dinov2_default_config = load_config("ssl_default_config")
+
+
+def load_and_merge_config(config_name: str):
+    default_config = OmegaConf.create(dinov2_default_config)
+    loaded_config = load_config(config_name)
+    return OmegaConf.merge(default_config, loaded_config)

torchhub/facebookresearch_dinov2_main/dinov2/configs/eval/vitb14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_base
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/eval/vitg14_pretrain.yaml

@@ -0,0 +1,7 @@
+student:
+  arch: vit_giant2
+  patch_size: 14
+  ffn_layer: swiglufused
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/eval/vitl14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_large
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/eval/vits14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_small
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/ssl_default_config.yaml

@@ -0,0 +1,115 @@
+MODEL:
+  WEIGHTS: ''
+compute_precision:
+  grad_scaler: true
+  teacher:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+  student:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+dino:
+  loss_weight: 1.0
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+  koleo_loss_weight: 0.1
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  separate_head: false
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+train:
+  batch_size_per_gpu: 64
+  dataset_path: ImageNet:split=TRAIN
+  output_dir: .
+  saveckp_freq: 20
+  seed: 0
+  num_workers: 10
+  OFFICIAL_EPOCH_LENGTH: 1250
+  cache_dataset: true
+  centering: "centering" # or "sinkhorn_knopp"
+student:
+  arch: vit_large
+  patch_size: 16
+  drop_path_rate: 0.3
+  layerscale: 1.0e-05
+  drop_path_uniform: true
+  pretrained_weights: ''
+  ffn_layer: "mlp"
+  block_chunks: 0
+  qkv_bias: true
+  proj_bias: true
+  ffn_bias: true
+teacher:
+  momentum_teacher: 0.992
+  final_momentum_teacher: 1
+  warmup_teacher_temp: 0.04
+  teacher_temp: 0.07
+  warmup_teacher_temp_epochs: 30
+optim:
+  epochs: 100
+  weight_decay: 0.04
+  weight_decay_end: 0.4
+  base_lr: 0.004  # learning rate for a batch size of 1024
+  lr: 0.  # will be set after applying scaling rule
+  warmup_epochs: 10
+  min_lr: 1.0e-06
+  clip_grad: 3.0
+  freeze_last_layer_epochs: 1
+  scaling_rule: sqrt_wrt_1024
+  patch_embed_lr_mult: 0.2
+  layerwise_decay: 0.9
+  adamw_beta1: 0.9
+  adamw_beta2: 0.999
+crops:
+  global_crops_scale:
+  - 0.32
+  - 1.0
+  local_crops_number: 8
+  local_crops_scale:
+  - 0.05
+  - 0.32
+  global_crops_size: 224
+  local_crops_size: 96
+evaluation:
+  eval_period_iterations: 12500

torchhub/facebookresearch_dinov2_main/dinov2/configs/train/vitg14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 12
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_giant2
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/train/vitl14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 32
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_large
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

torchhub/facebookresearch_dinov2_main/dinov2/configs/train/vitl16_short.yaml

@@ -0,0 +1,6 @@
+# this corresponds to the default config
+train:
+  dataset_path: ImageNet:split=TRAIN
+  batch_size_per_gpu: 64
+student:
+  block_chunks: 4

torchhub/facebookresearch_dinov2_main/dinov2/data/__init__.py

@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .adapters import DatasetWithEnumeratedTargets
+from .loaders import make_data_loader, make_dataset, SamplerType
+from .collate import collate_data_and_cast
+from .masking import MaskingGenerator
+from .augmentations import DataAugmentationDINO

torchhub/facebookresearch_dinov2_main/dinov2/data/adapters.py

@@ -0,0 +1,29 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Tuple
+
+from torch.utils.data import Dataset
+
+
+class DatasetWithEnumeratedTargets(Dataset):
+    def __init__(self, dataset):
+        self._dataset = dataset
+
+    def get_image_data(self, index: int) -> bytes:
+        return self._dataset.get_image_data(index)
+
+    def get_target(self, index: int) -> Tuple[Any, int]:
+        target = self._dataset.get_target(index)
+        return (index, target)
+
+    def __getitem__(self, index: int) -> Tuple[Any, Tuple[Any, int]]:
+        image, target = self._dataset[index]
+        target = index if target is None else target
+        return image, (index, target)
+
+    def __len__(self) -> int:
+        return len(self._dataset)

torchhub/facebookresearch_dinov2_main/dinov2/data/augmentations.py

@@ -0,0 +1,119 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from torchvision import transforms
+
+from .transforms import (
+    GaussianBlur,
+    make_normalize_transform,
+)
+
+
+logger = logging.getLogger("dinov2")
+
+
+class DataAugmentationDINO(object):
+    def __init__(
+        self,
+        global_crops_scale,
+        local_crops_scale,
+        local_crops_number,
+        global_crops_size=224,
+        local_crops_size=96,
+    ):
+        self.global_crops_scale = global_crops_scale
+        self.local_crops_scale = local_crops_scale
+        self.local_crops_number = local_crops_number
+        self.global_crops_size = global_crops_size
+        self.local_crops_size = local_crops_size
+
+        logger.info("###################################")
+        logger.info("Using data augmentation parameters:")
+        logger.info(f"global_crops_scale: {global_crops_scale}")
+        logger.info(f"local_crops_scale: {local_crops_scale}")
+        logger.info(f"local_crops_number: {local_crops_number}")
+        logger.info(f"global_crops_size: {global_crops_size}")
+        logger.info(f"local_crops_size: {local_crops_size}")
+        logger.info("###################################")
+
+        # random resized crop and flip
+        self.geometric_augmentation_global = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    global_crops_size, scale=global_crops_scale, interpolation=transforms.InterpolationMode.BICUBIC
+                ),
+                transforms.RandomHorizontalFlip(p=0.5),
+            ]
+        )
+
+        self.geometric_augmentation_local = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    local_crops_size, scale=local_crops_scale, interpolation=transforms.InterpolationMode.BICUBIC
+                ),
+                transforms.RandomHorizontalFlip(p=0.5),
+            ]
+        )
+
+        # color distorsions / blurring
+        color_jittering = transforms.Compose(
+            [
+                transforms.RandomApply(
+                    [transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)],
+                    p=0.8,
+                ),
+                transforms.RandomGrayscale(p=0.2),
+            ]
+        )
+
+        global_transfo1_extra = GaussianBlur(p=1.0)
+
+        global_transfo2_extra = transforms.Compose(
+            [
+                GaussianBlur(p=0.1),
+                transforms.RandomSolarize(threshold=128, p=0.2),
+            ]
+        )
+
+        local_transfo_extra = GaussianBlur(p=0.5)
+
+        # normalization
+        self.normalize = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                make_normalize_transform(),
+            ]
+        )
+
+        self.global_transfo1 = transforms.Compose([color_jittering, global_transfo1_extra, self.normalize])
+        self.global_transfo2 = transforms.Compose([color_jittering, global_transfo2_extra, self.normalize])
+        self.local_transfo = transforms.Compose([color_jittering, local_transfo_extra, self.normalize])
+
+    def __call__(self, image):
+        output = {}
+
+        # global crops:
+        im1_base = self.geometric_augmentation_global(image)
+        global_crop_1 = self.global_transfo1(im1_base)
+
+        im2_base = self.geometric_augmentation_global(image)
+        global_crop_2 = self.global_transfo2(im2_base)
+
+        output["global_crops"] = [global_crop_1, global_crop_2]
+
+        # global crops for teacher:
+        output["global_crops_teacher"] = [global_crop_1, global_crop_2]
+
+        # local crops:
+        local_crops = [
+            self.local_transfo(self.geometric_augmentation_local(image)) for _ in range(self.local_crops_number)
+        ]
+        output["local_crops"] = local_crops
+        output["offsets"] = ()
+
+        return output

torchhub/facebookresearch_dinov2_main/dinov2/data/collate.py

@@ -0,0 +1,50 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import random
+
+
+def collate_data_and_cast(samples_list, mask_ratio_tuple, mask_probability, dtype, n_tokens=None, mask_generator=None):
+    # dtype = torch.half  # TODO: Remove
+
+    n_global_crops = len(samples_list[0][0]["global_crops"])
+    n_local_crops = len(samples_list[0][0]["local_crops"])
+
+    collated_global_crops = torch.stack([s[0]["global_crops"][i] for i in range(n_global_crops) for s in samples_list])
+
+    collated_local_crops = torch.stack([s[0]["local_crops"][i] for i in range(n_local_crops) for s in samples_list])
+
+    B = len(collated_global_crops)
+    N = n_tokens
+    n_samples_masked = int(B * mask_probability)
+    probs = torch.linspace(*mask_ratio_tuple, n_samples_masked + 1)
+    upperbound = 0
+    masks_list = []
+    for i in range(0, n_samples_masked):
+        prob_min = probs[i]
+        prob_max = probs[i + 1]
+        masks_list.append(torch.BoolTensor(mask_generator(int(N * random.uniform(prob_min, prob_max)))))
+        upperbound += int(N * prob_max)
+    for i in range(n_samples_masked, B):
+        masks_list.append(torch.BoolTensor(mask_generator(0)))
+
+    random.shuffle(masks_list)
+
+    collated_masks = torch.stack(masks_list).flatten(1)
+    mask_indices_list = collated_masks.flatten().nonzero().flatten()
+
+    masks_weight = (1 / collated_masks.sum(-1).clamp(min=1.0)).unsqueeze(-1).expand_as(collated_masks)[collated_masks]
+
+    return {
+        "collated_global_crops": collated_global_crops.to(dtype),
+        "collated_local_crops": collated_local_crops.to(dtype),
+        "collated_masks": collated_masks,
+        "mask_indices_list": mask_indices_list,
+        "masks_weight": masks_weight,
+        "upperbound": upperbound,
+        "n_masked_patches": torch.full((1,), fill_value=mask_indices_list.shape[0], dtype=torch.long),
+    }

torchhub/facebookresearch_dinov2_main/dinov2/data/datasets/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .image_net import ImageNet
+from .image_net_22k import ImageNet22k

torchhub/facebookresearch_dinov2_main/dinov2/data/datasets/decoders.py

@@ -0,0 +1,32 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from io import BytesIO
+from typing import Any
+
+from PIL import Image
+
+
+class Decoder:
+    def decode(self) -> Any:
+        raise NotImplementedError
+
+
+class ImageDataDecoder(Decoder):
+    def __init__(self, image_data: bytes) -> None:
+        self._image_data = image_data
+
+    def decode(self) -> Image:
+        f = BytesIO(self._image_data)
+        return Image.open(f).convert(mode="RGB")
+
+
+class TargetDecoder(Decoder):
+    def __init__(self, target: Any):
+        self._target = target
+
+    def decode(self) -> Any:
+        return self._target

torchhub/facebookresearch_dinov2_main/dinov2/data/datasets/extended.py

@@ -0,0 +1,39 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, Tuple
+
+from torchvision.datasets import VisionDataset
+
+from .decoders import TargetDecoder, ImageDataDecoder
+
+
+class ExtendedVisionDataset(VisionDataset):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)  # type: ignore
+
+    def get_image_data(self, index: int) -> bytes:
+        raise NotImplementedError
+
+    def get_target(self, index: int) -> Any:
+        raise NotImplementedError
+
+    def __getitem__(self, index: int) -> Tuple[Any, Any]:
+        try:
+            image_data = self.get_image_data(index)
+            image = ImageDataDecoder(image_data).decode()
+        except Exception as e:
+            raise RuntimeError(f"can not read image for sample {index}") from e
+        target = self.get_target(index)
+        target = TargetDecoder(target).decode()
+
+        if self.transforms is not None:
+            image, target = self.transforms(image, target)
+
+        return image, target
+
+    def __len__(self) -> int:
+        raise NotImplementedError

torchhub/facebookresearch_dinov2_main/dinov2/data/datasets/image_net.py

@@ -0,0 +1,291 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import csv
+from enum import Enum
+import logging
+import os
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+
+from .extended import ExtendedVisionDataset
+
+
+logger = logging.getLogger("dinov2")
+_Target = int
+
+
+class _Split(Enum):
+    TRAIN = "train"
+    VAL = "val"
+    TEST = "test"  # NOTE: torchvision does not support the test split
+
+    @property
+    def length(self) -> int:
+        split_lengths = {
+            _Split.TRAIN: 1_281_167,
+            _Split.VAL: 50_000,
+            _Split.TEST: 100_000,
+        }
+        return split_lengths[self]
+
+    def get_dirname(self, class_id: Optional[str] = None) -> str:
+        return self.value if class_id is None else os.path.join(self.value, class_id)
+
+    def get_image_relpath(self, actual_index: int, class_id: Optional[str] = None) -> str:
+        dirname = self.get_dirname(class_id)
+        if self == _Split.TRAIN:
+            basename = f"{class_id}_{actual_index}"
+        else:  # self in (_Split.VAL, _Split.TEST):
+            basename = f"ILSVRC2012_{self.value}_{actual_index:08d}"
+        return os.path.join(dirname, basename + ".JPEG")
+
+    def parse_image_relpath(self, image_relpath: str) -> Tuple[str, int]:
+        assert self != _Split.TEST
+        dirname, filename = os.path.split(image_relpath)
+        class_id = os.path.split(dirname)[-1]
+        basename, _ = os.path.splitext(filename)
+        actual_index = int(basename.split("_")[-1])
+        return class_id, actual_index
+
+
+class ImageNet(ExtendedVisionDataset):
+    Target = Union[_Target]
+    Split = Union[_Split]
+
+    def __init__(
+        self,
+        *,
+        split: "ImageNet.Split",
+        root: str,
+        extra: str,
+        transforms: Optional[Callable] = None,
+        transform: Optional[Callable] = None,
+        target_transform: Optional[Callable] = None,
+    ) -> None:
+        super().__init__(root, transforms, transform, target_transform)
+        self._extra_root = extra
+        self._split = split
+
+        self._entries = None
+        self._class_ids = None
+        self._class_names = None
+
+    @property
+    def split(self) -> "ImageNet.Split":
+        return self._split
+
+    def _get_extra_full_path(self, extra_path: str) -> str:
+        return os.path.join(self._extra_root, extra_path)
+
+    def _load_extra(self, extra_path: str) -> np.ndarray:
+        extra_full_path = self._get_extra_full_path(extra_path)
+        return np.load(extra_full_path, mmap_mode="r")
+
+    def _save_extra(self, extra_array: np.ndarray, extra_path: str) -> None:
+        extra_full_path = self._get_extra_full_path(extra_path)
+        os.makedirs(self._extra_root, exist_ok=True)
+        np.save(extra_full_path, extra_array)
+
+    @property
+    def _entries_path(self) -> str:
+        return f"entries-{self._split.value.upper()}.npy"
+
+    @property
+    def _class_ids_path(self) -> str:
+        return f"class-ids-{self._split.value.upper()}.npy"
+
+    @property
+    def _class_names_path(self) -> str:
+        return f"class-names-{self._split.value.upper()}.npy"
+
+    def _get_entries(self) -> np.ndarray:
+        if self._entries is None:
+            self._entries = self._load_extra(self._entries_path)
+        assert self._entries is not None
+        return self._entries
+
+    def _get_class_ids(self) -> np.ndarray:
+        if self._split == _Split.TEST:
+            assert False, "Class IDs are not available in TEST split"
+        if self._class_ids is None:
+            self._class_ids = self._load_extra(self._class_ids_path)
+        assert self._class_ids is not None
+        return self._class_ids
+
+    def _get_class_names(self) -> np.ndarray:
+        if self._split == _Split.TEST:
+            assert False, "Class names are not available in TEST split"
+        if self._class_names is None:
+            self._class_names = self._load_extra(self._class_names_path)
+        assert self._class_names is not None
+        return self._class_names
+
+    def find_class_id(self, class_index: int) -> str:
+        class_ids = self._get_class_ids()
+        return str(class_ids[class_index])
+
+    def find_class_name(self, class_index: int) -> str:
+        class_names = self._get_class_names()
+        return str(class_names[class_index])
+
+    def get_image_data(self, index: int) -> bytes:
+        entries = self._get_entries()
+        actual_index = entries[index]["actual_index"]
+
+        class_id = self.get_class_id(index)
+
+        image_relpath = self.split.get_image_relpath(actual_index, class_id)
+        image_full_path = os.path.join(self.root, image_relpath)
+        with open(image_full_path, mode="rb") as f:
+            image_data = f.read()
+        return image_data
+
+    def get_target(self, index: int) -> Optional[Target]:
+        entries = self._get_entries()
+        class_index = entries[index]["class_index"]
+        return None if self.split == _Split.TEST else int(class_index)
+
+    def get_targets(self) -> Optional[np.ndarray]:
+        entries = self._get_entries()
+        return None if self.split == _Split.TEST else entries["class_index"]
+
+    def get_class_id(self, index: int) -> Optional[str]:
+        entries = self._get_entries()
+        class_id = entries[index]["class_id"]
+        return None if self.split == _Split.TEST else str(class_id)
+
+    def get_class_name(self, index: int) -> Optional[str]:
+        entries = self._get_entries()
+        class_name = entries[index]["class_name"]
+        return None if self.split == _Split.TEST else str(class_name)
+
+    def __len__(self) -> int:
+        entries = self._get_entries()
+        assert len(entries) == self.split.length
+        return len(entries)
+
+    def _load_labels(self, labels_path: str) -> List[Tuple[str, str]]:
+        labels_full_path = os.path.join(self.root, labels_path)
+        labels = []
+
+        try:
+            with open(labels_full_path, "r") as f:
+                reader = csv.reader(f)
+                for row in reader:
+                    class_id, class_name = row
+                    labels.append((class_id, class_name))
+        except OSError as e:
+            raise RuntimeError(f'can not read labels file "{labels_full_path}"') from e
+
+        return labels
+
+    def _dump_entries(self) -> None:
+        split = self.split
+        if split == ImageNet.Split.TEST:
+            dataset = None
+            sample_count = split.length
+            max_class_id_length, max_class_name_length = 0, 0
+        else:
+            labels_path = "labels.txt"
+            logger.info(f'loading labels from "{labels_path}"')
+            labels = self._load_labels(labels_path)
+
+            # NOTE: Using torchvision ImageFolder for consistency
+            from torchvision.datasets import ImageFolder
+
+            dataset_root = os.path.join(self.root, split.get_dirname())
+            dataset = ImageFolder(dataset_root)
+            sample_count = len(dataset)
+            max_class_id_length, max_class_name_length = -1, -1
+            for sample in dataset.samples:
+                _, class_index = sample
+                class_id, class_name = labels[class_index]
+                max_class_id_length = max(len(class_id), max_class_id_length)
+                max_class_name_length = max(len(class_name), max_class_name_length)
+
+        dtype = np.dtype(
+            [
+                ("actual_index", "<u4"),
+                ("class_index", "<u4"),
+                ("class_id", f"U{max_class_id_length}"),
+                ("class_name", f"U{max_class_name_length}"),
+            ]
+        )
+        entries_array = np.empty(sample_count, dtype=dtype)
+
+        if split == ImageNet.Split.TEST:
+            old_percent = -1
+            for index in range(sample_count):
+                percent = 100 * (index + 1) // sample_count
+                if percent > old_percent:
+                    logger.info(f"creating entries: {percent}%")
+                    old_percent = percent
+
+                actual_index = index + 1
+                class_index = np.uint32(-1)
+                class_id, class_name = "", ""
+                entries_array[index] = (actual_index, class_index, class_id, class_name)
+        else:
+            class_names = {class_id: class_name for class_id, class_name in labels}
+
+            assert dataset
+            old_percent = -1
+            for index in range(sample_count):
+                percent = 100 * (index + 1) // sample_count
+                if percent > old_percent:
+                    logger.info(f"creating entries: {percent}%")
+                    old_percent = percent
+
+                image_full_path, class_index = dataset.samples[index]
+                image_relpath = os.path.relpath(image_full_path, self.root)
+                class_id, actual_index = split.parse_image_relpath(image_relpath)
+                class_name = class_names[class_id]
+                entries_array[index] = (actual_index, class_index, class_id, class_name)
+
+        logger.info(f'saving entries to "{self._entries_path}"')
+        self._save_extra(entries_array, self._entries_path)
+
+    def _dump_class_ids_and_names(self) -> None:
+        split = self.split
+        if split == ImageNet.Split.TEST:
+            return
+
+        entries_array = self._load_extra(self._entries_path)
+
+        max_class_id_length, max_class_name_length, max_class_index = -1, -1, -1
+        for entry in entries_array:
+            class_index, class_id, class_name = (
+                entry["class_index"],
+                entry["class_id"],
+                entry["class_name"],
+            )
+            max_class_index = max(int(class_index), max_class_index)
+            max_class_id_length = max(len(str(class_id)), max_class_id_length)
+            max_class_name_length = max(len(str(class_name)), max_class_name_length)
+
+        class_count = max_class_index + 1
+        class_ids_array = np.empty(class_count, dtype=f"U{max_class_id_length}")
+        class_names_array = np.empty(class_count, dtype=f"U{max_class_name_length}")
+        for entry in entries_array:
+            class_index, class_id, class_name = (
+                entry["class_index"],
+                entry["class_id"],
+                entry["class_name"],
+            )
+            class_ids_array[class_index] = class_id
+            class_names_array[class_index] = class_name
+
+        logger.info(f'saving class IDs to "{self._class_ids_path}"')
+        self._save_extra(class_ids_array, self._class_ids_path)
+
+        logger.info(f'saving class names to "{self._class_names_path}"')
+        self._save_extra(class_names_array, self._class_names_path)
+
+    def dump_extra(self) -> None:
+        self._dump_entries()
+        self._dump_class_ids_and_names()

torchhub/facebookresearch_dinov2_main/dinov2/data/datasets/image_net_22k.py

@@ -0,0 +1,303 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from dataclasses import dataclass
+from enum import Enum
+from functools import lru_cache
+from gzip import GzipFile
+from io import BytesIO
+from mmap import ACCESS_READ, mmap
+import os
+from typing import Any, Callable, List, Optional, Set, Tuple
+import warnings
+
+import numpy as np
+
+from .extended import ExtendedVisionDataset
+
+
+_Labels = int
+
+_DEFAULT_MMAP_CACHE_SIZE = 16  # Warning: This can exhaust file descriptors
+
+
+@dataclass
+class _ClassEntry:
+    block_offset: int
+    maybe_filename: Optional[str] = None
+
+
+@dataclass
+class _Entry:
+    class_index: int  # noqa: E701
+    start_offset: int
+    end_offset: int
+    filename: str
+
+
+class _Split(Enum):
+    TRAIN = "train"
+    VAL = "val"
+
+    @property
+    def length(self) -> int:
+        return {
+            _Split.TRAIN: 11_797_647,
+            _Split.VAL: 561_050,
+        }[self]
+
+    def entries_path(self):
+        return f"imagenet21kp_{self.value}.txt"
+
+
+def _get_tarball_path(class_id: str) -> str:
+    return f"{class_id}.tar"
+
+
+def _make_mmap_tarball(tarballs_root: str, mmap_cache_size: int):
+    @lru_cache(maxsize=mmap_cache_size)
+    def _mmap_tarball(class_id: str) -> mmap:
+        tarball_path = _get_tarball_path(class_id)
+        tarball_full_path = os.path.join(tarballs_root, tarball_path)
+        with open(tarball_full_path) as f:
+            return mmap(fileno=f.fileno(), length=0, access=ACCESS_READ)
+
+    return _mmap_tarball
+
+
+class ImageNet22k(ExtendedVisionDataset):
+    _GZIPPED_INDICES: Set[int] = {
+        841_545,
+        1_304_131,
+        2_437_921,
+        2_672_079,
+        2_795_676,
+        2_969_786,
+        6_902_965,
+        6_903_550,
+        6_903_628,
+        7_432_557,
+        7_432_589,
+        7_813_809,
+        8_329_633,
+        10_296_990,
+        10_417_652,
+        10_492_265,
+        10_598_078,
+        10_782_398,
+        10_902_612,
+        11_203_736,
+        11_342_890,
+        11_397_596,
+        11_589_762,
+        11_705_103,
+        12_936_875,
+        13_289_782,
+    }
+    Labels = _Labels
+
+    def __init__(
+        self,
+        *,
+        root: str,
+        extra: str,
+        transforms: Optional[Callable] = None,
+        transform: Optional[Callable] = None,
+        target_transform: Optional[Callable] = None,
+        mmap_cache_size: int = _DEFAULT_MMAP_CACHE_SIZE,
+    ) -> None:
+        super().__init__(root, transforms, transform, target_transform)
+        self._extra_root = extra
+
+        entries_path = self._get_entries_path(root)
+        self._entries = self._load_extra(entries_path)
+
+        class_ids_path = self._get_class_ids_path(root)
+        self._class_ids = self._load_extra(class_ids_path)
+
+        self._gzipped_indices = ImageNet22k._GZIPPED_INDICES
+        self._mmap_tarball = _make_mmap_tarball(self._tarballs_root, mmap_cache_size)
+
+    def _get_entries_path(self, root: Optional[str] = None) -> str:
+        return "entries.npy"
+
+    def _get_class_ids_path(self, root: Optional[str] = None) -> str:
+        return "class-ids.npy"
+
+    def _find_class_ids(self, path: str) -> List[str]:
+        class_ids = []
+
+        with os.scandir(path) as entries:
+            for entry in entries:
+                root, ext = os.path.splitext(entry.name)
+                if ext != ".tar":
+                    continue
+                class_ids.append(root)
+
+        return sorted(class_ids)
+
+    def _load_entries_class_ids(self, root: Optional[str] = None) -> Tuple[List[_Entry], List[str]]:
+        root = self.get_root(root)
+        entries: List[_Entry] = []
+        class_ids = self._find_class_ids(root)
+
+        for class_index, class_id in enumerate(class_ids):
+            path = os.path.join(root, "blocks", f"{class_id}.log")
+            class_entries = []
+
+            try:
+                with open(path) as f:
+                    for line in f:
+                        line = line.rstrip()
+                        block, filename = line.split(":")
+                        block_offset = int(block[6:])
+                        filename = filename[1:]
+
+                        maybe_filename = None
+                        if filename != "** Block of NULs **":
+                            maybe_filename = filename
+                            _, ext = os.path.splitext(filename)
+                            # assert ext == ".JPEG"
+
+                        class_entry = _ClassEntry(block_offset, maybe_filename)
+                        class_entries.append(class_entry)
+            except OSError as e:
+                raise RuntimeError(f'can not read blocks file "{path}"') from e
+
+            assert class_entries[-1].maybe_filename is None
+
+            for class_entry1, class_entry2 in zip(class_entries, class_entries[1:]):
+                assert class_entry1.block_offset <= class_entry2.block_offset
+                start_offset = 512 * class_entry1.block_offset
+                end_offset = 512 * class_entry2.block_offset
+                assert class_entry1.maybe_filename is not None
+                filename = class_entry1.maybe_filename
+                entry = _Entry(class_index, start_offset, end_offset, filename)
+                # Skip invalid image files (PIL throws UnidentifiedImageError)
+                if filename == "n06470073_47249.JPEG":
+                    continue
+                entries.append(entry)
+
+        return entries, class_ids
+
+    def _load_extra(self, extra_path: str) -> np.ndarray:
+        extra_root = self._extra_root
+        extra_full_path = os.path.join(extra_root, extra_path)
+        return np.load(extra_full_path, mmap_mode="r")
+
+    def _save_extra(self, extra_array: np.ndarray, extra_path: str) -> None:
+        extra_root = self._extra_root
+        extra_full_path = os.path.join(extra_root, extra_path)
+        os.makedirs(extra_root, exist_ok=True)
+        np.save(extra_full_path, extra_array)
+
+    @property
+    def _tarballs_root(self) -> str:
+        return self.root
+
+    def find_class_id(self, class_index: int) -> str:
+        return str(self._class_ids[class_index])
+
+    def get_image_data(self, index: int) -> bytes:
+        entry = self._entries[index]
+        class_id = entry["class_id"]
+        class_mmap = self._mmap_tarball(class_id)
+
+        start_offset, end_offset = entry["start_offset"], entry["end_offset"]
+        try:
+            mapped_data = class_mmap[start_offset:end_offset]
+            data = mapped_data[512:]  # Skip entry header block
+
+            if len(data) >= 2 and tuple(data[:2]) == (0x1F, 0x8B):
+                assert index in self._gzipped_indices, f"unexpected gzip header for sample {index}"
+                with GzipFile(fileobj=BytesIO(data)) as g:
+                    data = g.read()
+        except Exception as e:
+            raise RuntimeError(f"can not retrieve image data for sample {index} " f'from "{class_id}" tarball') from e
+
+        return data
+
+    def get_target(self, index: int) -> Any:
+        return int(self._entries[index]["class_index"])
+
+    def get_targets(self) -> np.ndarray:
+        return self._entries["class_index"]
+
+    def get_class_id(self, index: int) -> str:
+        return str(self._entries[index]["class_id"])
+
+    def get_class_ids(self) -> np.ndarray:
+        return self._entries["class_id"]
+
+    def __getitem__(self, index: int) -> Tuple[Any, Any]:
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            return super().__getitem__(index)
+
+    def __len__(self) -> int:
+        return len(self._entries)
+
+    def _dump_entries(self, *args, **kwargs) -> None:
+        entries, class_ids = self._load_entries_class_ids(*args, **kwargs)
+
+        max_class_id_length, max_filename_length, max_class_index = -1, -1, -1
+        for entry in entries:
+            class_id = class_ids[entry.class_index]
+            max_class_index = max(entry.class_index, max_class_index)
+            max_class_id_length = max(len(class_id), max_class_id_length)
+            max_filename_length = max(len(entry.filename), max_filename_length)
+
+        dtype = np.dtype(
+            [
+                ("class_index", "<u4"),
+                ("class_id", f"U{max_class_id_length}"),
+                ("start_offset", "<u4"),
+                ("end_offset", "<u4"),
+                ("filename", f"U{max_filename_length}"),
+            ]
+        )
+        sample_count = len(entries)
+        entries_array = np.empty(sample_count, dtype=dtype)
+        for i, entry in enumerate(entries):
+            class_index = entry.class_index
+            class_id = class_ids[class_index]
+            start_offset = entry.start_offset
+            end_offset = entry.end_offset
+            filename = entry.filename
+            entries_array[i] = (
+                class_index,
+                class_id,
+                start_offset,
+                end_offset,
+                filename,
+            )
+
+        entries_path = self._get_entries_path(*args, **kwargs)
+        self._save_extra(entries_array, entries_path)
+
+    def _dump_class_ids(self, *args, **kwargs) -> None:
+        entries_path = self._get_entries_path(*args, **kwargs)
+        entries_array = self._load_extra(entries_path)
+
+        max_class_id_length, max_class_index = -1, -1
+        for entry in entries_array:
+            class_index, class_id = entry["class_index"], entry["class_id"]
+            max_class_index = max(int(class_index), max_class_index)
+            max_class_id_length = max(len(str(class_id)), max_class_id_length)
+
+        class_ids_array = np.empty(max_class_index + 1, dtype=f"U{max_class_id_length}")
+        for entry in entries_array:
+            class_index, class_id = entry["class_index"], entry["class_id"]
+            class_ids_array[class_index] = class_id
+        class_ids_path = self._get_class_ids_path(*args, **kwargs)
+        self._save_extra(class_ids_array, class_ids_path)
+
+    def _dump_extra(self, *args, **kwargs) -> None:
+        self._dump_entries(*args, *kwargs)
+        self._dump_class_ids(*args, *kwargs)
+
+    def dump_extra(self, root: Optional[str] = None) -> None:
+        return self._dump_extra(root)

torchhub/facebookresearch_dinov2_main/dinov2/data/loaders.py

@@ -0,0 +1,223 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from enum import Enum
+from typing import Any, Callable, List, Optional, TypeVar
+
+import torch
+from torch.utils.data import Sampler
+
+from .datasets import ImageNet, ImageNet22k
+from .samplers import EpochSampler, InfiniteSampler, ShardedInfiniteSampler
+
+
+logger = logging.getLogger("dinov2")
+
+
+class SamplerType(Enum):
+    DISTRIBUTED = 0
+    EPOCH = 1
+    INFINITE = 2
+    SHARDED_INFINITE = 3
+    SHARDED_INFINITE_NEW = 4
+
+
+def _make_bool_str(b: bool) -> str:
+    return "yes" if b else "no"
+
+
+def _make_sample_transform(image_transform: Optional[Callable] = None, target_transform: Optional[Callable] = None):
+    def transform(sample):
+        image, target = sample
+        if image_transform is not None:
+            image = image_transform(image)
+        if target_transform is not None:
+            target = target_transform(target)
+        return image, target
+
+    return transform
+
+
+def _parse_dataset_str(dataset_str: str):
+    tokens = dataset_str.split(":")
+
+    name = tokens[0]
+    kwargs = {}
+
+    for token in tokens[1:]:
+        key, value = token.split("=")
+        assert key in ("root", "extra", "split")
+        kwargs[key] = value
+
+    if name == "ImageNet":
+        class_ = ImageNet
+        if "split" in kwargs:
+            kwargs["split"] = ImageNet.Split[kwargs["split"]]
+    elif name == "ImageNet22k":
+        class_ = ImageNet22k
+    else:
+        raise ValueError(f'Unsupported dataset "{name}"')
+
+    return class_, kwargs
+
+
+def make_dataset(
+    *,
+    dataset_str: str,
+    transform: Optional[Callable] = None,
+    target_transform: Optional[Callable] = None,
+):
+    """
+    Creates a dataset with the specified parameters.
+
+    Args:
+        dataset_str: A dataset string description (e.g. ImageNet:split=TRAIN).
+        transform: A transform to apply to images.
+        target_transform: A transform to apply to targets.
+
+    Returns:
+        The created dataset.
+    """
+    logger.info(f'using dataset: "{dataset_str}"')
+
+    class_, kwargs = _parse_dataset_str(dataset_str)
+    dataset = class_(transform=transform, target_transform=target_transform, **kwargs)
+
+    logger.info(f"# of dataset samples: {len(dataset):,d}")
+
+    # Aggregated datasets do not expose (yet) these attributes, so add them.
+    if not hasattr(dataset, "transform"):
+        setattr(dataset, "transform", transform)
+    if not hasattr(dataset, "target_transform"):
+        setattr(dataset, "target_transform", target_transform)
+
+    return dataset
+
+
+def _make_sampler(
+    *,
+    dataset,
+    type: Optional[SamplerType] = None,
+    shuffle: bool = False,
+    seed: int = 0,
+    size: int = -1,
+    advance: int = 0,
+) -> Optional[Sampler]:
+    sample_count = len(dataset)
+
+    if type == SamplerType.INFINITE:
+        logger.info("sampler: infinite")
+        if size > 0:
+            raise ValueError("sampler size > 0 is invalid")
+        return InfiniteSampler(
+            sample_count=sample_count,
+            shuffle=shuffle,
+            seed=seed,
+            advance=advance,
+        )
+    elif type in (SamplerType.SHARDED_INFINITE, SamplerType.SHARDED_INFINITE_NEW):
+        logger.info("sampler: sharded infinite")
+        if size > 0:
+            raise ValueError("sampler size > 0 is invalid")
+        # TODO: Remove support for old shuffling
+        use_new_shuffle_tensor_slice = type == SamplerType.SHARDED_INFINITE_NEW
+        return ShardedInfiniteSampler(
+            sample_count=sample_count,
+            shuffle=shuffle,
+            seed=seed,
+            advance=advance,
+            use_new_shuffle_tensor_slice=use_new_shuffle_tensor_slice,
+        )
+    elif type == SamplerType.EPOCH:
+        logger.info("sampler: epoch")
+        if advance > 0:
+            raise NotImplementedError("sampler advance > 0 is not supported")
+        size = size if size > 0 else sample_count
+        logger.info(f"# of samples / epoch: {size:,d}")
+        return EpochSampler(
+            size=size,
+            sample_count=sample_count,
+            shuffle=shuffle,
+            seed=seed,
+        )
+    elif type == SamplerType.DISTRIBUTED:
+        logger.info("sampler: distributed")
+        if size > 0:
+            raise ValueError("sampler size > 0 is invalid")
+        if advance > 0:
+            raise ValueError("sampler advance > 0 is invalid")
+        return torch.utils.data.DistributedSampler(
+            dataset=dataset,
+            shuffle=shuffle,
+            seed=seed,
+            drop_last=False,
+        )
+
+    logger.info("sampler: none")
+    return None
+
+
+T = TypeVar("T")
+
+
+def make_data_loader(
+    *,
+    dataset,
+    batch_size: int,
+    num_workers: int,
+    shuffle: bool = True,
+    seed: int = 0,
+    sampler_type: Optional[SamplerType] = SamplerType.INFINITE,
+    sampler_size: int = -1,
+    sampler_advance: int = 0,
+    drop_last: bool = True,
+    persistent_workers: bool = False,
+    collate_fn: Optional[Callable[[List[T]], Any]] = None,
+):
+    """
+    Creates a data loader with the specified parameters.
+
+    Args:
+        dataset: A dataset (third party, LaViDa or WebDataset).
+        batch_size: The size of batches to generate.
+        num_workers: The number of workers to use.
+        shuffle: Whether to shuffle samples.
+        seed: The random seed to use.
+        sampler_type: Which sampler to use: EPOCH, INFINITE, SHARDED_INFINITE, SHARDED_INFINITE_NEW, DISTRIBUTED or None.
+        sampler_size: The number of images per epoch (when applicable) or -1 for the entire dataset.
+        sampler_advance: How many samples to skip (when applicable).
+        drop_last: Whether the last non-full batch of data should be dropped.
+        persistent_workers: maintain the workers Dataset instances alive after a dataset has been consumed once.
+        collate_fn: Function that performs batch collation
+    """
+
+    sampler = _make_sampler(
+        dataset=dataset,
+        type=sampler_type,
+        shuffle=shuffle,
+        seed=seed,
+        size=sampler_size,
+        advance=sampler_advance,
+    )
+
+    logger.info("using PyTorch data loader")
+    data_loader = torch.utils.data.DataLoader(
+        dataset,
+        sampler=sampler,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        pin_memory=True,
+        drop_last=drop_last,
+        persistent_workers=persistent_workers,
+        collate_fn=collate_fn,
+    )
+
+    try:
+        logger.info(f"# of batches: {len(data_loader):,d}")
+    except TypeError:  # data loader has no length
+        logger.info("infinite data loader")
+    return data_loader

torchhub/facebookresearch_dinov2_main/dinov2/data/masking.py

@@ -0,0 +1,87 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+import math
+import numpy as np
+
+
+class MaskingGenerator:
+    def __init__(
+        self,
+        input_size,
+        num_masking_patches=None,
+        min_num_patches=4,
+        max_num_patches=None,
+        min_aspect=0.3,
+        max_aspect=None,
+    ):
+        if not isinstance(input_size, tuple):
+            input_size = (input_size,) * 2
+        self.height, self.width = input_size
+
+        self.num_patches = self.height * self.width
+        self.num_masking_patches = num_masking_patches
+
+        self.min_num_patches = min_num_patches
+        self.max_num_patches = num_masking_patches if max_num_patches is None else max_num_patches
+
+        max_aspect = max_aspect or 1 / min_aspect
+        self.log_aspect_ratio = (math.log(min_aspect), math.log(max_aspect))
+
+    def __repr__(self):
+        repr_str = "Generator(%d, %d -> [%d ~ %d], max = %d, %.3f ~ %.3f)" % (
+            self.height,
+            self.width,
+            self.min_num_patches,
+            self.max_num_patches,
+            self.num_masking_patches,
+            self.log_aspect_ratio[0],
+            self.log_aspect_ratio[1],
+        )
+        return repr_str
+
+    def get_shape(self):
+        return self.height, self.width
+
+    def _mask(self, mask, max_mask_patches):
+        delta = 0
+        for _ in range(10):
+            target_area = random.uniform(self.min_num_patches, max_mask_patches)
+            aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
+            h = int(round(math.sqrt(target_area * aspect_ratio)))
+            w = int(round(math.sqrt(target_area / aspect_ratio)))
+            if w < self.width and h < self.height:
+                top = random.randint(0, self.height - h)
+                left = random.randint(0, self.width - w)
+
+                num_masked = mask[top : top + h, left : left + w].sum()
+                # Overlap
+                if 0 < h * w - num_masked <= max_mask_patches:
+                    for i in range(top, top + h):
+                        for j in range(left, left + w):
+                            if mask[i, j] == 0:
+                                mask[i, j] = 1
+                                delta += 1
+
+                if delta > 0:
+                    break
+        return delta
+
+    def __call__(self, num_masking_patches=0):
+        mask = np.zeros(shape=self.get_shape(), dtype=bool)
+        mask_count = 0
+        while mask_count < num_masking_patches:
+            max_mask_patches = num_masking_patches - mask_count
+            max_mask_patches = min(max_mask_patches, self.max_num_patches)
+
+            delta = self._mask(mask, max_mask_patches)
+            if delta == 0:
+                break
+            else:
+                mask_count += delta
+
+        return mask

torchhub/facebookresearch_dinov2_main/dinov2/data/samplers.py

@@ -0,0 +1,230 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import itertools
+from typing import Any, Optional
+import warnings
+
+import numpy as np
+import torch
+from torch.utils.data.sampler import Sampler
+
+import dinov2.distributed as distributed
+
+
+class EpochSampler(Sampler):
+    def __init__(
+        self,
+        *,
+        size: int,
+        sample_count: int,
+        shuffle: bool = False,
+        seed: int = 0,
+        start: Optional[int] = None,
+        step: Optional[int] = None,
+    ):
+        self._size = size
+        self._sample_count = sample_count
+        self._shuffle = shuffle
+        self._seed = seed
+        self._start = distributed.get_global_rank() if start is None else start
+        self._step = distributed.get_global_size() if step is None else step
+        self._epoch = 0
+
+    def __iter__(self):
+        count = (self._size + self._sample_count - 1) // self._sample_count
+        tiled_indices = np.tile(np.arange(self._sample_count), count)
+        if self._shuffle:
+            seed = self._seed * self._epoch if self._seed != 0 else self._epoch
+            rng = np.random.default_rng(seed)
+            iterable = rng.choice(tiled_indices, self._size, replace=False)
+        else:
+            iterable = tiled_indices[: self._size]
+
+        yield from itertools.islice(iterable, self._start, None, self._step)
+
+    def __len__(self):
+        return (self._size - self._start + self._step - 1) // self._step
+
+    def set_epoch(self, epoch):
+        self._epoch = epoch
+
+
+def _get_numpy_dtype(size: int) -> Any:
+    return np.int32 if size <= 2**31 else np.int64
+
+
+def _get_torch_dtype(size: int) -> Any:
+    return torch.int32 if size <= 2**31 else torch.int64
+
+
+def _generate_randperm_indices(*, size: int, generator: torch.Generator):
+    """Generate the indices of a random permutation."""
+    dtype = _get_torch_dtype(size)
+    # This is actually matching PyTorch's CPU implementation, see: https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/TensorFactories.cpp#L900-L921
+    perm = torch.arange(size, dtype=dtype)
+    for i in range(size):
+        j = torch.randint(i, size, size=(1,), generator=generator).item()
+
+        # Always swap even if no-op
+        value = perm[j].item()
+        perm[j] = perm[i].item()
+        perm[i] = value
+        yield value
+
+
+class InfiniteSampler(Sampler):
+    def __init__(
+        self,
+        *,
+        sample_count: int,
+        shuffle: bool = False,
+        seed: int = 0,
+        start: Optional[int] = None,
+        step: Optional[int] = None,
+        advance: int = 0,
+    ):
+        self._sample_count = sample_count
+        self._seed = seed
+        self._shuffle = shuffle
+        self._start = distributed.get_global_rank() if start is None else start
+        self._step = distributed.get_global_size() if step is None else step
+        self._advance = advance
+
+    def __iter__(self):
+        if self._shuffle:
+            iterator = self._shuffled_iterator()
+        else:
+            iterator = self._iterator()
+
+        yield from itertools.islice(iterator, self._advance, None)
+
+    def _iterator(self):
+        assert not self._shuffle
+
+        while True:
+            iterable = range(self._sample_count)
+            yield from itertools.islice(iterable, self._start, None, self._step)
+
+    def _shuffled_iterator(self):
+        assert self._shuffle
+
+        # Instantiate a generator here (rather than in the ctor) to keep the class
+        # picklable (requirement of mp.spawn)
+        generator = torch.Generator().manual_seed(self._seed)
+
+        while True:
+            iterable = _generate_randperm_indices(size=self._sample_count, generator=generator)
+            yield from itertools.islice(iterable, self._start, None, self._step)
+
+
+# The following function is somewhat equivalent to _new_shuffle_tensor_slice below,
+# but avoids a full in-place random permutation generation.
+def _shuffle_tensor_slice(
+    *, tensor: torch.Tensor, start: int = 0, step: int = 1, generator: torch.Generator
+) -> np.ndarray:
+    stop = len(tensor)
+    count = stop // step
+    drop_count = stop - step * count
+    if drop_count:
+        warnings.warn(f"# of dropped samples: {drop_count}")
+
+    dtype = _get_numpy_dtype(stop)
+    result = np.empty(count, dtype=dtype)
+
+    for i in range(count):
+        j = torch.randint(0, i + 1, size=(1,), generator=generator).item() if i > 0 else 0
+
+        result[i] = result[j]
+        result[j] = tensor[start + i * step].item()
+
+    return result
+
+
+def _new_shuffle_tensor_slice(
+    *, tensor: torch.Tensor, start: int = 0, step: int = 1, generator: torch.Generator
+) -> np.ndarray:
+    stop = len(tensor)
+    count = stop // step
+    dtype = torch.int64  # Needed for using randperm result as indices
+    count = stop // step
+    drop_count = stop - step * count
+    if drop_count:
+        warnings.warn(f"# of dropped samples: {drop_count}")
+    indices = torch.randperm(count, dtype=dtype, generator=generator)
+    return tensor[start::step][indices].numpy()
+
+
+def _make_seed(seed: int, start: int, iter_count: int) -> int:
+    # NOTE: Tried a few variants (including iter_count << 32), this one worked best.
+    return seed + start + (iter_count << 24)
+
+
+class ShardedInfiniteSampler(Sampler):
+    def __init__(
+        self,
+        *,
+        sample_count: int,
+        shuffle: bool = False,
+        seed: int = 0,
+        start: Optional[int] = None,
+        step: Optional[int] = None,
+        advance: int = 0,
+        use_new_shuffle_tensor_slice: bool = False,
+    ):
+        self._sample_count = sample_count
+        self._seed = seed
+        self._shuffle = shuffle
+        self._start = distributed.get_global_rank() if start is None else start
+        self._step = distributed.get_global_size() if step is None else step
+        self._advance = advance
+        self._iter_count = 0
+        self._shuffle_tensor_slice_fn = (
+            _new_shuffle_tensor_slice if use_new_shuffle_tensor_slice else _shuffle_tensor_slice
+        )
+
+    def __iter__(self):
+        iter_count = self._advance // self._sample_count
+        if iter_count > 0:
+            self._advance -= iter_count * self._sample_count
+            self._iter_count += iter_count
+
+        if self._shuffle:
+            iterator = self._shuffled_iterator()
+        else:
+            iterator = self._iterator()
+
+        yield from itertools.islice(iterator, self._advance, None)
+
+    def _iterator(self):
+        assert not self._shuffle
+
+        while True:
+            iterable = range(self._sample_count)
+            yield from itertools.islice(iterable, self._start, None, self._step)
+
+    def _shuffled_iterator(self):
+        assert self._shuffle
+
+        # Instantiate a generator here (rather than in the ctor) to be keep the class
+        # picklable (requirement of mp.spawn)
+        generator = torch.Generator()
+
+        # Always shuffle everything first
+        generator.manual_seed(self._seed)
+        dtype = _get_torch_dtype(self._sample_count)
+        perm = torch.randperm(self._sample_count, dtype=dtype, generator=generator)
+
+        while True:
+            # Re-seed on each iteration to allow skipping whole permutations
+            seed = _make_seed(self._seed, self._start, self._iter_count)
+            generator.manual_seed(seed)
+
+            iterable = self._shuffle_tensor_slice_fn(
+                tensor=perm, start=self._start, step=self._step, generator=generator
+            )
+            yield from iterable
+            self._iter_count += 1

torchhub/facebookresearch_dinov2_main/dinov2/data/transforms.py

@@ -0,0 +1,92 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Sequence
+
+import torch
+from torchvision import transforms
+
+
+class GaussianBlur(transforms.RandomApply):
+    """
+    Apply Gaussian Blur to the PIL image.
+    """
+
+    def __init__(self, *, p: float = 0.5, radius_min: float = 0.1, radius_max: float = 2.0):
+        # NOTE: torchvision is applying 1 - probability to return the original image
+        keep_p = 1 - p
+        transform = transforms.GaussianBlur(kernel_size=9, sigma=(radius_min, radius_max))
+        super().__init__(transforms=[transform], p=keep_p)
+
+
+class MaybeToTensor(transforms.ToTensor):
+    """
+    Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor, or keep as is if already a tensor.
+    """
+
+    def __call__(self, pic):
+        """
+        Args:
+            pic (PIL Image, numpy.ndarray or torch.tensor): Image to be converted to tensor.
+        Returns:
+            Tensor: Converted image.
+        """
+        if isinstance(pic, torch.Tensor):
+            return pic
+        return super().__call__(pic)
+
+
+# Use timm's names
+IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
+
+
+def make_normalize_transform(
+    mean: Sequence[float] = IMAGENET_DEFAULT_MEAN,
+    std: Sequence[float] = IMAGENET_DEFAULT_STD,
+) -> transforms.Normalize:
+    return transforms.Normalize(mean=mean, std=std)
+
+
+# This roughly matches torchvision's preset for classification training:
+#   https://github.com/pytorch/vision/blob/main/references/classification/presets.py#L6-L44
+def make_classification_train_transform(
+    *,
+    crop_size: int = 224,
+    interpolation=transforms.InterpolationMode.BICUBIC,
+    hflip_prob: float = 0.5,
+    mean: Sequence[float] = IMAGENET_DEFAULT_MEAN,
+    std: Sequence[float] = IMAGENET_DEFAULT_STD,
+):
+    transforms_list = [transforms.RandomResizedCrop(crop_size, interpolation=interpolation)]
+    if hflip_prob > 0.0:
+        transforms_list.append(transforms.RandomHorizontalFlip(hflip_prob))
+    transforms_list.extend(
+        [
+            MaybeToTensor(),
+            make_normalize_transform(mean=mean, std=std),
+        ]
+    )
+    return transforms.Compose(transforms_list)
+
+
+# This matches (roughly) torchvision's preset for classification evaluation:
+#   https://github.com/pytorch/vision/blob/main/references/classification/presets.py#L47-L69
+def make_classification_eval_transform(
+    *,
+    resize_size: int = 256,
+    interpolation=transforms.InterpolationMode.BICUBIC,
+    crop_size: int = 224,
+    mean: Sequence[float] = IMAGENET_DEFAULT_MEAN,
+    std: Sequence[float] = IMAGENET_DEFAULT_STD,
+) -> transforms.Compose:
+    transforms_list = [
+        transforms.Resize(resize_size, interpolation=interpolation),
+        transforms.CenterCrop(crop_size),
+        MaybeToTensor(),
+        make_normalize_transform(mean=mean, std=std),
+    ]
+    return transforms.Compose(transforms_list)

torchhub/facebookresearch_dinov2_main/dinov2/distributed/__init__.py

@@ -0,0 +1,271 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import random
+import re
+import socket
+from typing import Dict, List
+
+import torch
+import torch.distributed as dist
+
+_LOCAL_RANK = -1
+_LOCAL_WORLD_SIZE = -1
+
+
+def is_enabled() -> bool:
+    """
+    Returns:
+        True if distributed training is enabled
+    """
+    return dist.is_available() and dist.is_initialized()
+
+
+def get_global_size() -> int:
+    """
+    Returns:
+        The number of processes in the process group
+    """
+    return dist.get_world_size() if is_enabled() else 1
+
+
+def get_global_rank() -> int:
+    """
+    Returns:
+        The rank of the current process within the global process group.
+    """
+    return dist.get_rank() if is_enabled() else 0
+
+
+def get_local_rank() -> int:
+    """
+    Returns:
+        The rank of the current process within the local (per-machine) process group.
+    """
+    if not is_enabled():
+        return 0
+    assert 0 <= _LOCAL_RANK < _LOCAL_WORLD_SIZE
+    return _LOCAL_RANK
+
+
+def get_local_size() -> int:
+    """
+    Returns:
+        The size of the per-machine process group,
+        i.e. the number of processes per machine.
+    """
+    if not is_enabled():
+        return 1
+    assert 0 <= _LOCAL_RANK < _LOCAL_WORLD_SIZE
+    return _LOCAL_WORLD_SIZE
+
+
+def is_main_process() -> bool:
+    """
+    Returns:
+        True if the current process is the main one.
+    """
+    return get_global_rank() == 0
+
+
+def _restrict_print_to_main_process() -> None:
+    """
+    This function disables printing when not in the main process
+    """
+    import builtins as __builtin__
+
+    builtin_print = __builtin__.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop("force", False)
+        if is_main_process() or force:
+            builtin_print(*args, **kwargs)
+
+    __builtin__.print = print
+
+
+def _get_master_port(seed: int = 0) -> int:
+    MIN_MASTER_PORT, MAX_MASTER_PORT = (20_000, 60_000)
+
+    master_port_str = os.environ.get("MASTER_PORT")
+    if master_port_str is None:
+        rng = random.Random(seed)
+        return rng.randint(MIN_MASTER_PORT, MAX_MASTER_PORT)
+
+    return int(master_port_str)
+
+
+def _get_available_port() -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        # A "" host address means INADDR_ANY i.e. binding to all interfaces.
+        # Note this is not compatible with IPv6.
+        s.bind(("", 0))
+        port = s.getsockname()[1]
+        return port
+
+
+_TORCH_DISTRIBUTED_ENV_VARS = (
+    "MASTER_ADDR",
+    "MASTER_PORT",
+    "RANK",
+    "WORLD_SIZE",
+    "LOCAL_RANK",
+    "LOCAL_WORLD_SIZE",
+)
+
+
+def _collect_env_vars() -> Dict[str, str]:
+    return {env_var: os.environ[env_var] for env_var in _TORCH_DISTRIBUTED_ENV_VARS if env_var in os.environ}
+
+
+def _is_slurm_job_process() -> bool:
+    return "SLURM_JOB_ID" in os.environ
+
+
+def _parse_slurm_node_list(s: str) -> List[str]:
+    nodes = []
+    # Extract "hostname", "hostname[1-2,3,4-5]," substrings
+    p = re.compile(r"(([^\[]+)(?:\[([^\]]+)\])?),?")
+    for m in p.finditer(s):
+        prefix, suffixes = s[m.start(2) : m.end(2)], s[m.start(3) : m.end(3)]
+        for suffix in suffixes.split(","):
+            span = suffix.split("-")
+            if len(span) == 1:
+                nodes.append(prefix + suffix)
+            else:
+                width = len(span[0])
+                start, end = int(span[0]), int(span[1]) + 1
+                nodes.extend([prefix + f"{i:0{width}}" for i in range(start, end)])
+    return nodes
+
+
+def _check_env_variable(key: str, new_value: str):
+    # Only check for difference with preset environment variables
+    if key in os.environ and os.environ[key] != new_value:
+        raise RuntimeError(f"Cannot export environment variables as {key} is already set")
+
+
+class _TorchDistributedEnvironment:
+    def __init__(self):
+        self.master_addr = "127.0.0.1"
+        self.master_port = 0
+        self.rank = -1
+        self.world_size = -1
+        self.local_rank = -1
+        self.local_world_size = -1
+
+        if _is_slurm_job_process():
+            return self._set_from_slurm_env()
+
+        env_vars = _collect_env_vars()
+        if not env_vars:
+            # Environment is not set
+            pass
+        elif len(env_vars) == len(_TORCH_DISTRIBUTED_ENV_VARS):
+            # Environment is fully set
+            return self._set_from_preset_env()
+        else:
+            # Environment is partially set
+            collected_env_vars = ", ".join(env_vars.keys())
+            raise RuntimeError(f"Partially set environment: {collected_env_vars}")
+
+        if torch.cuda.device_count() > 0:
+            return self._set_from_local()
+
+        raise RuntimeError("Can't initialize PyTorch distributed environment")
+
+    # Slurm job created with sbatch, submitit, etc...
+    def _set_from_slurm_env(self):
+        # logger.info("Initialization from Slurm environment")
+        job_id = int(os.environ["SLURM_JOB_ID"])
+        node_count = int(os.environ["SLURM_JOB_NUM_NODES"])
+        nodes = _parse_slurm_node_list(os.environ["SLURM_JOB_NODELIST"])
+        assert len(nodes) == node_count
+
+        self.master_addr = nodes[0]
+        self.master_port = _get_master_port(seed=job_id)
+        self.rank = int(os.environ["SLURM_PROCID"])
+        self.world_size = int(os.environ["SLURM_NTASKS"])
+        assert self.rank < self.world_size
+        self.local_rank = int(os.environ["SLURM_LOCALID"])
+        self.local_world_size = self.world_size // node_count
+        assert self.local_rank < self.local_world_size
+
+    # Single node job with preset environment (i.e. torchrun)
+    def _set_from_preset_env(self):
+        # logger.info("Initialization from preset environment")
+        self.master_addr = os.environ["MASTER_ADDR"]
+        self.master_port = os.environ["MASTER_PORT"]
+        self.rank = int(os.environ["RANK"])
+        self.world_size = int(os.environ["WORLD_SIZE"])
+        assert self.rank < self.world_size
+        self.local_rank = int(os.environ["LOCAL_RANK"])
+        self.local_world_size = int(os.environ["LOCAL_WORLD_SIZE"])
+        assert self.local_rank < self.local_world_size
+
+    # Single node and GPU job (i.e. local script run)
+    def _set_from_local(self):
+        # logger.info("Initialization from local")
+        self.master_addr = "127.0.0.1"
+        self.master_port = _get_available_port()
+        self.rank = 0
+        self.world_size = 1
+        self.local_rank = 0
+        self.local_world_size = 1
+
+    def export(self, *, overwrite: bool) -> "_TorchDistributedEnvironment":
+        # See the "Environment variable initialization" section from
+        # https://pytorch.org/docs/stable/distributed.html for the complete list of
+        # environment variables required for the env:// initialization method.
+        env_vars = {
+            "MASTER_ADDR": self.master_addr,
+            "MASTER_PORT": str(self.master_port),
+            "RANK": str(self.rank),
+            "WORLD_SIZE": str(self.world_size),
+            "LOCAL_RANK": str(self.local_rank),
+            "LOCAL_WORLD_SIZE": str(self.local_world_size),
+        }
+        if not overwrite:
+            for k, v in env_vars.items():
+                _check_env_variable(k, v)
+
+        os.environ.update(env_vars)
+        return self
+
+
+def enable(*, set_cuda_current_device: bool = True, overwrite: bool = False, allow_nccl_timeout: bool = False):
+    """Enable distributed mode
+
+    Args:
+        set_cuda_current_device: If True, call torch.cuda.set_device() to set the
+            current PyTorch CUDA device to the one matching the local rank.
+        overwrite: If True, overwrites already set variables. Else fails.
+    """
+
+    global _LOCAL_RANK, _LOCAL_WORLD_SIZE
+    if _LOCAL_RANK >= 0 or _LOCAL_WORLD_SIZE >= 0:
+        raise RuntimeError("Distributed mode has already been enabled")
+    torch_env = _TorchDistributedEnvironment()
+    torch_env.export(overwrite=overwrite)
+
+    if set_cuda_current_device:
+        torch.cuda.set_device(torch_env.local_rank)
+
+    if allow_nccl_timeout:
+        # This allows to use torch distributed timeout in a NCCL backend
+        key, value = "NCCL_ASYNC_ERROR_HANDLING", "1"
+        if not overwrite:
+            _check_env_variable(key, value)
+        os.environ[key] = value
+
+    dist.init_process_group(backend="nccl")
+    dist.barrier()
+
+    # Finalize setup
+    _LOCAL_RANK = torch_env.local_rank
+    _LOCAL_WORLD_SIZE = torch_env.local_world_size
+    _restrict_print_to_main_process()

torchhub/facebookresearch_dinov2_main/dinov2/eval/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torchhub/facebookresearch_dinov2_main/dinov2/eval/knn.py

@@ -0,0 +1,405 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+from functools import partial
+import json
+import logging
+import os
+import sys
+from typing import List, Optional
+
+import torch
+from torch.nn.functional import one_hot, softmax
+
+import dinov2.distributed as distributed
+from dinov2.data import SamplerType, make_data_loader, make_dataset
+from dinov2.data.transforms import make_classification_eval_transform
+from dinov2.eval.metrics import AccuracyAveraging, build_topk_accuracy_metric
+from dinov2.eval.setup import get_args_parser as get_setup_args_parser
+from dinov2.eval.setup import setup_and_build_model
+from dinov2.eval.utils import ModelWithNormalize, evaluate, extract_features
+
+
+logger = logging.getLogger("dinov2")
+
+
+def get_args_parser(
+    description: Optional[str] = None,
+    parents: Optional[List[argparse.ArgumentParser]] = None,
+    add_help: bool = True,
+):
+    parents = parents or []
+    setup_args_parser = get_setup_args_parser(parents=parents, add_help=False)
+    parents = [setup_args_parser]
+    parser = argparse.ArgumentParser(
+        description=description,
+        parents=parents,
+        add_help=add_help,
+    )
+    parser.add_argument(
+        "--train-dataset",
+        dest="train_dataset_str",
+        type=str,
+        help="Training dataset",
+    )
+    parser.add_argument(
+        "--val-dataset",
+        dest="val_dataset_str",
+        type=str,
+        help="Validation dataset",
+    )
+    parser.add_argument(
+        "--nb_knn",
+        nargs="+",
+        type=int,
+        help="Number of NN to use. 20 is usually working the best.",
+    )
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        help="Temperature used in the voting coefficient",
+    )
+    parser.add_argument(
+        "--gather-on-cpu",
+        action="store_true",
+        help="Whether to gather the train features on cpu, slower"
+        "but useful to avoid OOM for large datasets (e.g. ImageNet22k).",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        help="Batch size.",
+    )
+    parser.add_argument(
+        "--n-per-class-list",
+        nargs="+",
+        type=int,
+        help="Number to take per class",
+    )
+    parser.add_argument(
+        "--n-tries",
+        type=int,
+        help="Number of tries",
+    )
+    parser.set_defaults(
+        train_dataset_str="ImageNet:split=TRAIN",
+        val_dataset_str="ImageNet:split=VAL",
+        nb_knn=[10, 20, 100, 200],
+        temperature=0.07,
+        batch_size=256,
+        n_per_class_list=[-1],
+        n_tries=1,
+    )
+    return parser
+
+
+class KnnModule(torch.nn.Module):
+    """
+    Gets knn of test features from all processes on a chunk of the train features
+
+    Each rank gets a chunk of the train features as well as a chunk of the test features.
+    In `compute_neighbors`, for each rank one after the other, its chunk of test features
+    is sent to all devices, partial knns are computed with each chunk of train features
+    then collated back on the original device.
+    """
+
+    def __init__(self, train_features, train_labels, nb_knn, T, device, num_classes=1000):
+        super().__init__()
+
+        self.global_rank = distributed.get_global_rank()
+        self.global_size = distributed.get_global_size()
+
+        self.device = device
+        self.train_features_rank_T = train_features.chunk(self.global_size)[self.global_rank].T.to(self.device)
+        self.candidates = train_labels.chunk(self.global_size)[self.global_rank].view(1, -1).to(self.device)
+
+        self.nb_knn = nb_knn
+        self.max_k = max(self.nb_knn)
+        self.T = T
+        self.num_classes = num_classes
+
+    def _get_knn_sims_and_labels(self, similarity, train_labels):
+        topk_sims, indices = similarity.topk(self.max_k, largest=True, sorted=True)
+        neighbors_labels = torch.gather(train_labels, 1, indices)
+        return topk_sims, neighbors_labels
+
+    def _similarity_for_rank(self, features_rank, source_rank):
+        # Send the features from `source_rank` to all ranks
+        broadcast_shape = torch.tensor(features_rank.shape).to(self.device)
+        torch.distributed.broadcast(broadcast_shape, source_rank)
+
+        broadcasted = features_rank
+        if self.global_rank != source_rank:
+            broadcasted = torch.zeros(*broadcast_shape, dtype=features_rank.dtype, device=self.device)
+        torch.distributed.broadcast(broadcasted, source_rank)
+
+        # Compute the neighbors for `source_rank` among `train_features_rank_T`
+        similarity_rank = torch.mm(broadcasted, self.train_features_rank_T)
+        candidate_labels = self.candidates.expand(len(similarity_rank), -1)
+        return self._get_knn_sims_and_labels(similarity_rank, candidate_labels)
+
+    def _gather_all_knn_for_rank(self, topk_sims, neighbors_labels, target_rank):
+        # Gather all neighbors for `target_rank`
+        topk_sims_rank = retrieved_rank = None
+        if self.global_rank == target_rank:
+            topk_sims_rank = [torch.zeros_like(topk_sims) for _ in range(self.global_size)]
+            retrieved_rank = [torch.zeros_like(neighbors_labels) for _ in range(self.global_size)]
+
+        torch.distributed.gather(topk_sims, topk_sims_rank, dst=target_rank)
+        torch.distributed.gather(neighbors_labels, retrieved_rank, dst=target_rank)
+
+        if self.global_rank == target_rank:
+            # Perform a second top-k on the k * global_size retrieved neighbors
+            topk_sims_rank = torch.cat(topk_sims_rank, dim=1)
+            retrieved_rank = torch.cat(retrieved_rank, dim=1)
+            results = self._get_knn_sims_and_labels(topk_sims_rank, retrieved_rank)
+            return results
+        return None
+
+    def compute_neighbors(self, features_rank):
+        for rank in range(self.global_size):
+            topk_sims, neighbors_labels = self._similarity_for_rank(features_rank, rank)
+            results = self._gather_all_knn_for_rank(topk_sims, neighbors_labels, rank)
+            if results is not None:
+                topk_sims_rank, neighbors_labels_rank = results
+        return topk_sims_rank, neighbors_labels_rank
+
+    def forward(self, features_rank):
+        """
+        Compute the results on all values of `self.nb_knn` neighbors from the full `self.max_k`
+        """
+        assert all(k <= self.max_k for k in self.nb_knn)
+
+        topk_sims, neighbors_labels = self.compute_neighbors(features_rank)
+        batch_size = neighbors_labels.shape[0]
+        topk_sims_transform = softmax(topk_sims / self.T, 1)
+        matmul = torch.mul(
+            one_hot(neighbors_labels, num_classes=self.num_classes),
+            topk_sims_transform.view(batch_size, -1, 1),
+        )
+        probas_for_k = {k: torch.sum(matmul[:, :k, :], 1) for k in self.nb_knn}
+        return probas_for_k
+
+
+class DictKeysModule(torch.nn.Module):
+    def __init__(self, keys):
+        super().__init__()
+        self.keys = keys
+
+    def forward(self, features_dict, targets):
+        for k in self.keys:
+            features_dict = features_dict[k]
+        return {"preds": features_dict, "target": targets}
+
+
+def create_module_dict(*, module, n_per_class_list, n_tries, nb_knn, train_features, train_labels):
+    modules = {}
+    mapping = create_class_indices_mapping(train_labels)
+    for npc in n_per_class_list:
+        if npc < 0:  # Only one try needed when using the full data
+            full_module = module(
+                train_features=train_features,
+                train_labels=train_labels,
+                nb_knn=nb_knn,
+            )
+            modules["full"] = ModuleDictWithForward({"1": full_module})
+            continue
+        all_tries = {}
+        for t in range(n_tries):
+            final_indices = filter_train(mapping, npc, seed=t)
+            k_list = list(set(nb_knn + [npc]))
+            k_list = sorted([el for el in k_list if el <= npc])
+            all_tries[str(t)] = module(
+                train_features=train_features[final_indices],
+                train_labels=train_labels[final_indices],
+                nb_knn=k_list,
+            )
+        modules[f"{npc} per class"] = ModuleDictWithForward(all_tries)
+
+    return ModuleDictWithForward(modules)
+
+
+def filter_train(mapping, n_per_class, seed):
+    torch.manual_seed(seed)
+    final_indices = []
+    for k in mapping.keys():
+        index = torch.randperm(len(mapping[k]))[:n_per_class]
+        final_indices.append(mapping[k][index])
+    return torch.cat(final_indices).squeeze()
+
+
+def create_class_indices_mapping(labels):
+    unique_labels, inverse = torch.unique(labels, return_inverse=True)
+    mapping = {unique_labels[i]: (inverse == i).nonzero() for i in range(len(unique_labels))}
+    return mapping
+
+
+class ModuleDictWithForward(torch.nn.ModuleDict):
+    def forward(self, *args, **kwargs):
+        return {k: module(*args, **kwargs) for k, module in self._modules.items()}
+
+
+def eval_knn(
+    model,
+    train_dataset,
+    val_dataset,
+    accuracy_averaging,
+    nb_knn,
+    temperature,
+    batch_size,
+    num_workers,
+    gather_on_cpu,
+    n_per_class_list=[-1],
+    n_tries=1,
+):
+    model = ModelWithNormalize(model)
+
+    logger.info("Extracting features for train set...")
+    train_features, train_labels = extract_features(
+        model, train_dataset, batch_size, num_workers, gather_on_cpu=gather_on_cpu
+    )
+    logger.info(f"Train features created, shape {train_features.shape}.")
+
+    val_dataloader = make_data_loader(
+        dataset=val_dataset,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        sampler_type=SamplerType.DISTRIBUTED,
+        drop_last=False,
+        shuffle=False,
+        persistent_workers=True,
+    )
+    num_classes = train_labels.max() + 1
+    metric_collection = build_topk_accuracy_metric(accuracy_averaging, num_classes=num_classes)
+
+    device = torch.cuda.current_device()
+    partial_module = partial(KnnModule, T=temperature, device=device, num_classes=num_classes)
+    knn_module_dict = create_module_dict(
+        module=partial_module,
+        n_per_class_list=n_per_class_list,
+        n_tries=n_tries,
+        nb_knn=nb_knn,
+        train_features=train_features,
+        train_labels=train_labels,
+    )
+    postprocessors, metrics = {}, {}
+    for n_per_class, knn_module in knn_module_dict.items():
+        for t, knn_try in knn_module.items():
+            postprocessors = {
+                **postprocessors,
+                **{(n_per_class, t, k): DictKeysModule([n_per_class, t, k]) for k in knn_try.nb_knn},
+            }
+            metrics = {**metrics, **{(n_per_class, t, k): metric_collection.clone() for k in knn_try.nb_knn}}
+    model_with_knn = torch.nn.Sequential(model, knn_module_dict)
+
+    # ============ evaluation ... ============
+    logger.info("Start the k-NN classification.")
+    _, results_dict = evaluate(model_with_knn, val_dataloader, postprocessors, metrics, device)
+
+    # Averaging the results over the n tries for each value of n_per_class
+    for n_per_class, knn_module in knn_module_dict.items():
+        first_try = list(knn_module.keys())[0]
+        k_list = knn_module[first_try].nb_knn
+        for k in k_list:
+            keys = results_dict[(n_per_class, first_try, k)].keys()  # keys are e.g. `top-1` and `top-5`
+            results_dict[(n_per_class, k)] = {
+                key: torch.mean(torch.stack([results_dict[(n_per_class, t, k)][key] for t in knn_module.keys()]))
+                for key in keys
+            }
+            for t in knn_module.keys():
+                del results_dict[(n_per_class, t, k)]
+
+    return results_dict
+
+
+def eval_knn_with_model(
+    model,
+    output_dir,
+    train_dataset_str="ImageNet:split=TRAIN",
+    val_dataset_str="ImageNet:split=VAL",
+    nb_knn=(10, 20, 100, 200),
+    temperature=0.07,
+    autocast_dtype=torch.float,
+    accuracy_averaging=AccuracyAveraging.MEAN_ACCURACY,
+    transform=None,
+    gather_on_cpu=False,
+    batch_size=256,
+    num_workers=5,
+    n_per_class_list=[-1],
+    n_tries=1,
+):
+    transform = transform or make_classification_eval_transform()
+
+    train_dataset = make_dataset(
+        dataset_str=train_dataset_str,
+        transform=transform,
+    )
+    val_dataset = make_dataset(
+        dataset_str=val_dataset_str,
+        transform=transform,
+    )
+
+    with torch.cuda.amp.autocast(dtype=autocast_dtype):
+        results_dict_knn = eval_knn(
+            model=model,
+            train_dataset=train_dataset,
+            val_dataset=val_dataset,
+            accuracy_averaging=accuracy_averaging,
+            nb_knn=nb_knn,
+            temperature=temperature,
+            batch_size=batch_size,
+            num_workers=num_workers,
+            gather_on_cpu=gather_on_cpu,
+            n_per_class_list=n_per_class_list,
+            n_tries=n_tries,
+        )
+
+    results_dict = {}
+    if distributed.is_main_process():
+        for knn_ in results_dict_knn.keys():
+            top1 = results_dict_knn[knn_]["top-1"].item() * 100.0
+            top5 = results_dict_knn[knn_]["top-5"].item() * 100.0
+            results_dict[f"{knn_} Top 1"] = top1
+            results_dict[f"{knn_} Top 5"] = top5
+            logger.info(f"{knn_} classifier result: Top1: {top1:.2f} Top5: {top5:.2f}")
+
+    metrics_file_path = os.path.join(output_dir, "results_eval_knn.json")
+    with open(metrics_file_path, "a") as f:
+        for k, v in results_dict.items():
+            f.write(json.dumps({k: v}) + "\n")
+
+    if distributed.is_enabled():
+        torch.distributed.barrier()
+    return results_dict
+
+
+def main(args):
+    model, autocast_dtype = setup_and_build_model(args)
+    eval_knn_with_model(
+        model=model,
+        output_dir=args.output_dir,
+        train_dataset_str=args.train_dataset_str,
+        val_dataset_str=args.val_dataset_str,
+        nb_knn=args.nb_knn,
+        temperature=args.temperature,
+        autocast_dtype=autocast_dtype,
+        accuracy_averaging=AccuracyAveraging.MEAN_ACCURACY,
+        transform=None,
+        gather_on_cpu=args.gather_on_cpu,
+        batch_size=args.batch_size,
+        num_workers=5,
+        n_per_class_list=args.n_per_class_list,
+        n_tries=args.n_tries,
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    description = "DINOv2 k-NN evaluation"
+    args_parser = get_args_parser(description=description)
+    args = args_parser.parse_args()
+    sys.exit(main(args))

torchhub/facebookresearch_dinov2_main/dinov2/eval/linear.py

@@ -0,0 +1,626 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+from functools import partial
+import json
+import logging
+import os
+import sys
+from typing import List, Optional
+
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn.parallel import DistributedDataParallel
+from fvcore.common.checkpoint import Checkpointer, PeriodicCheckpointer
+
+from dinov2.data import SamplerType, make_data_loader, make_dataset
+from dinov2.data.transforms import make_classification_eval_transform, make_classification_train_transform
+import dinov2.distributed as distributed
+from dinov2.eval.metrics import MetricType, build_metric
+from dinov2.eval.setup import get_args_parser as get_setup_args_parser
+from dinov2.eval.setup import setup_and_build_model
+from dinov2.eval.utils import ModelWithIntermediateLayers, evaluate
+from dinov2.logging import MetricLogger
+
+
+logger = logging.getLogger("dinov2")
+
+
+def get_args_parser(
+    description: Optional[str] = None,
+    parents: Optional[List[argparse.ArgumentParser]] = None,
+    add_help: bool = True,
+):
+    parents = parents or []
+    setup_args_parser = get_setup_args_parser(parents=parents, add_help=False)
+    parents = [setup_args_parser]
+    parser = argparse.ArgumentParser(
+        description=description,
+        parents=parents,
+        add_help=add_help,
+    )
+    parser.add_argument(
+        "--train-dataset",
+        dest="train_dataset_str",
+        type=str,
+        help="Training dataset",
+    )
+    parser.add_argument(
+        "--val-dataset",
+        dest="val_dataset_str",
+        type=str,
+        help="Validation dataset",
+    )
+    parser.add_argument(
+        "--test-datasets",
+        dest="test_dataset_strs",
+        type=str,
+        nargs="+",
+        help="Test datasets, none to reuse the validation dataset",
+    )
+    parser.add_argument(
+        "--epochs",
+        type=int,
+        help="Number of training epochs",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        help="Batch Size (per GPU)",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        help="Number de Workers",
+    )
+    parser.add_argument(
+        "--epoch-length",
+        type=int,
+        help="Length of an epoch in number of iterations",
+    )
+    parser.add_argument(
+        "--save-checkpoint-frequency",
+        type=int,
+        help="Number of epochs between two named checkpoint saves.",
+    )
+    parser.add_argument(
+        "--eval-period-iterations",
+        type=int,
+        help="Number of iterations between two evaluations.",
+    )
+    parser.add_argument(
+        "--learning-rates",
+        nargs="+",
+        type=float,
+        help="Learning rates to grid search.",
+    )
+    parser.add_argument(
+        "--no-resume",
+        action="store_true",
+        help="Whether to not resume from existing checkpoints",
+    )
+    parser.add_argument(
+        "--val-metric-type",
+        type=MetricType,
+        choices=list(MetricType),
+        help="Validation metric",
+    )
+    parser.add_argument(
+        "--test-metric-types",
+        type=MetricType,
+        choices=list(MetricType),
+        nargs="+",
+        help="Evaluation metric",
+    )
+    parser.add_argument(
+        "--classifier-fpath",
+        type=str,
+        help="Path to a file containing pretrained linear classifiers",
+    )
+    parser.add_argument(
+        "--val-class-mapping-fpath",
+        type=str,
+        help="Path to a file containing a mapping to adjust classifier outputs",
+    )
+    parser.add_argument(
+        "--test-class-mapping-fpaths",
+        nargs="+",
+        type=str,
+        help="Path to a file containing a mapping to adjust classifier outputs",
+    )
+    parser.set_defaults(
+        train_dataset_str="ImageNet:split=TRAIN",
+        val_dataset_str="ImageNet:split=VAL",
+        test_dataset_strs=None,
+        epochs=10,
+        batch_size=128,
+        num_workers=8,
+        epoch_length=1250,
+        save_checkpoint_frequency=20,
+        eval_period_iterations=1250,
+        learning_rates=[1e-5, 2e-5, 5e-5, 1e-4, 2e-4, 5e-4, 1e-3, 2e-3, 5e-3, 1e-2, 2e-2, 5e-2, 0.1],
+        val_metric_type=MetricType.MEAN_ACCURACY,
+        test_metric_types=None,
+        classifier_fpath=None,
+        val_class_mapping_fpath=None,
+        test_class_mapping_fpaths=[None],
+    )
+    return parser
+
+
+def has_ddp_wrapper(m: nn.Module) -> bool:
+    return isinstance(m, DistributedDataParallel)
+
+
+def remove_ddp_wrapper(m: nn.Module) -> nn.Module:
+    return m.module if has_ddp_wrapper(m) else m
+
+
+def _pad_and_collate(batch):
+    maxlen = max(len(targets) for image, targets in batch)
+    padded_batch = [
+        (image, np.pad(targets, (0, maxlen - len(targets)), constant_values=-1)) for image, targets in batch
+    ]
+    return torch.utils.data.default_collate(padded_batch)
+
+
+def create_linear_input(x_tokens_list, use_n_blocks, use_avgpool):
+    intermediate_output = x_tokens_list[-use_n_blocks:]
+    output = torch.cat([class_token for _, class_token in intermediate_output], dim=-1)
+    if use_avgpool:
+        output = torch.cat(
+            (
+                output,
+                torch.mean(intermediate_output[-1][0], dim=1),  # patch tokens
+            ),
+            dim=-1,
+        )
+        output = output.reshape(output.shape[0], -1)
+    return output.float()
+
+
+class LinearClassifier(nn.Module):
+    """Linear layer to train on top of frozen features"""
+
+    def __init__(self, out_dim, use_n_blocks, use_avgpool, num_classes=1000):
+        super().__init__()
+        self.out_dim = out_dim
+        self.use_n_blocks = use_n_blocks
+        self.use_avgpool = use_avgpool
+        self.num_classes = num_classes
+        self.linear = nn.Linear(out_dim, num_classes)
+        self.linear.weight.data.normal_(mean=0.0, std=0.01)
+        self.linear.bias.data.zero_()
+
+    def forward(self, x_tokens_list):
+        output = create_linear_input(x_tokens_list, self.use_n_blocks, self.use_avgpool)
+        return self.linear(output)
+
+
+class AllClassifiers(nn.Module):
+    def __init__(self, classifiers_dict):
+        super().__init__()
+        self.classifiers_dict = nn.ModuleDict()
+        self.classifiers_dict.update(classifiers_dict)
+
+    def forward(self, inputs):
+        return {k: v.forward(inputs) for k, v in self.classifiers_dict.items()}
+
+    def __len__(self):
+        return len(self.classifiers_dict)
+
+
+class LinearPostprocessor(nn.Module):
+    def __init__(self, linear_classifier, class_mapping=None):
+        super().__init__()
+        self.linear_classifier = linear_classifier
+        self.register_buffer("class_mapping", None if class_mapping is None else torch.LongTensor(class_mapping))
+
+    def forward(self, samples, targets):
+        preds = self.linear_classifier(samples)
+        return {
+            "preds": preds[:, self.class_mapping] if self.class_mapping is not None else preds,
+            "target": targets,
+        }
+
+
+def scale_lr(learning_rates, batch_size):
+    return learning_rates * (batch_size * distributed.get_global_size()) / 256.0
+
+
+def setup_linear_classifiers(sample_output, n_last_blocks_list, learning_rates, batch_size, num_classes=1000):
+    linear_classifiers_dict = nn.ModuleDict()
+    optim_param_groups = []
+    for n in n_last_blocks_list:
+        for avgpool in [False, True]:
+            for _lr in learning_rates:
+                lr = scale_lr(_lr, batch_size)
+                out_dim = create_linear_input(sample_output, use_n_blocks=n, use_avgpool=avgpool).shape[1]
+                linear_classifier = LinearClassifier(
+                    out_dim, use_n_blocks=n, use_avgpool=avgpool, num_classes=num_classes
+                )
+                linear_classifier = linear_classifier.cuda()
+                linear_classifiers_dict[
+                    f"classifier_{n}_blocks_avgpool_{avgpool}_lr_{lr:.5f}".replace(".", "_")
+                ] = linear_classifier
+                optim_param_groups.append({"params": linear_classifier.parameters(), "lr": lr})
+
+    linear_classifiers = AllClassifiers(linear_classifiers_dict)
+    if distributed.is_enabled():
+        linear_classifiers = nn.parallel.DistributedDataParallel(linear_classifiers)
+
+    return linear_classifiers, optim_param_groups
+
+
+@torch.no_grad()
+def evaluate_linear_classifiers(
+    feature_model,
+    linear_classifiers,
+    data_loader,
+    metric_type,
+    metrics_file_path,
+    training_num_classes,
+    iteration,
+    prefixstring="",
+    class_mapping=None,
+    best_classifier_on_val=None,
+):
+    logger.info("running validation !")
+
+    num_classes = len(class_mapping) if class_mapping is not None else training_num_classes
+    metric = build_metric(metric_type, num_classes=num_classes)
+    postprocessors = {k: LinearPostprocessor(v, class_mapping) for k, v in linear_classifiers.classifiers_dict.items()}
+    metrics = {k: metric.clone() for k in linear_classifiers.classifiers_dict}
+
+    _, results_dict_temp = evaluate(
+        feature_model,
+        data_loader,
+        postprocessors,
+        metrics,
+        torch.cuda.current_device(),
+    )
+
+    logger.info("")
+    results_dict = {}
+    max_accuracy = 0
+    best_classifier = ""
+    for i, (classifier_string, metric) in enumerate(results_dict_temp.items()):
+        logger.info(f"{prefixstring} -- Classifier: {classifier_string} * {metric}")
+        if (
+            best_classifier_on_val is None and metric["top-1"].item() > max_accuracy
+        ) or classifier_string == best_classifier_on_val:
+            max_accuracy = metric["top-1"].item()
+            best_classifier = classifier_string
+
+    results_dict["best_classifier"] = {"name": best_classifier, "accuracy": max_accuracy}
+
+    logger.info(f"best classifier: {results_dict['best_classifier']}")
+
+    if distributed.is_main_process():
+        with open(metrics_file_path, "a") as f:
+            f.write(f"iter: {iteration}\n")
+            for k, v in results_dict.items():
+                f.write(json.dumps({k: v}) + "\n")
+            f.write("\n")
+
+    return results_dict
+
+
+def eval_linear(
+    *,
+    feature_model,
+    linear_classifiers,
+    train_data_loader,
+    val_data_loader,
+    metrics_file_path,
+    optimizer,
+    scheduler,
+    output_dir,
+    max_iter,
+    checkpoint_period,  # In number of iter, creates a new file every period
+    running_checkpoint_period,  # Period to update main checkpoint file
+    eval_period,
+    metric_type,
+    training_num_classes,
+    resume=True,
+    classifier_fpath=None,
+    val_class_mapping=None,
+):
+    checkpointer = Checkpointer(linear_classifiers, output_dir, optimizer=optimizer, scheduler=scheduler)
+    start_iter = checkpointer.resume_or_load(classifier_fpath or "", resume=resume).get("iteration", -1) + 1
+
+    periodic_checkpointer = PeriodicCheckpointer(checkpointer, checkpoint_period, max_iter=max_iter)
+    iteration = start_iter
+    logger.info("Starting training from iteration {}".format(start_iter))
+    metric_logger = MetricLogger(delimiter="  ")
+    header = "Training"
+
+    for data, labels in metric_logger.log_every(
+        train_data_loader,
+        10,
+        header,
+        max_iter,
+        start_iter,
+    ):
+        data = data.cuda(non_blocking=True)
+        labels = labels.cuda(non_blocking=True)
+
+        features = feature_model(data)
+        outputs = linear_classifiers(features)
+
+        losses = {f"loss_{k}": nn.CrossEntropyLoss()(v, labels) for k, v in outputs.items()}
+        loss = sum(losses.values())
+
+        # compute the gradients
+        optimizer.zero_grad()
+        loss.backward()
+
+        # step
+        optimizer.step()
+        scheduler.step()
+
+        # log
+        if iteration % 10 == 0:
+            torch.cuda.synchronize()
+            metric_logger.update(loss=loss.item())
+            metric_logger.update(lr=optimizer.param_groups[0]["lr"])
+            print("lr", optimizer.param_groups[0]["lr"])
+
+        if iteration - start_iter > 5:
+            if iteration % running_checkpoint_period == 0:
+                torch.cuda.synchronize()
+                if distributed.is_main_process():
+                    logger.info("Checkpointing running_checkpoint")
+                    periodic_checkpointer.save("running_checkpoint_linear_eval", iteration=iteration)
+                torch.cuda.synchronize()
+        periodic_checkpointer.step(iteration)
+
+        if eval_period > 0 and (iteration + 1) % eval_period == 0 and iteration != max_iter - 1:
+            _ = evaluate_linear_classifiers(
+                feature_model=feature_model,
+                linear_classifiers=remove_ddp_wrapper(linear_classifiers),
+                data_loader=val_data_loader,
+                metrics_file_path=metrics_file_path,
+                prefixstring=f"ITER: {iteration}",
+                metric_type=metric_type,
+                training_num_classes=training_num_classes,
+                iteration=iteration,
+                class_mapping=val_class_mapping,
+            )
+            torch.cuda.synchronize()
+
+        iteration = iteration + 1
+
+    val_results_dict = evaluate_linear_classifiers(
+        feature_model=feature_model,
+        linear_classifiers=remove_ddp_wrapper(linear_classifiers),
+        data_loader=val_data_loader,
+        metrics_file_path=metrics_file_path,
+        metric_type=metric_type,
+        training_num_classes=training_num_classes,
+        iteration=iteration,
+        class_mapping=val_class_mapping,
+    )
+    return val_results_dict, feature_model, linear_classifiers, iteration
+
+
+def make_eval_data_loader(test_dataset_str, batch_size, num_workers, metric_type):
+    test_dataset = make_dataset(
+        dataset_str=test_dataset_str,
+        transform=make_classification_eval_transform(),
+    )
+    test_data_loader = make_data_loader(
+        dataset=test_dataset,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        sampler_type=SamplerType.DISTRIBUTED,
+        drop_last=False,
+        shuffle=False,
+        persistent_workers=False,
+        collate_fn=_pad_and_collate if metric_type == MetricType.IMAGENET_REAL_ACCURACY else None,
+    )
+    return test_data_loader
+
+
+def test_on_datasets(
+    feature_model,
+    linear_classifiers,
+    test_dataset_strs,
+    batch_size,
+    num_workers,
+    test_metric_types,
+    metrics_file_path,
+    training_num_classes,
+    iteration,
+    best_classifier_on_val,
+    prefixstring="",
+    test_class_mappings=[None],
+):
+    results_dict = {}
+    for test_dataset_str, class_mapping, metric_type in zip(test_dataset_strs, test_class_mappings, test_metric_types):
+        logger.info(f"Testing on {test_dataset_str}")
+        test_data_loader = make_eval_data_loader(test_dataset_str, batch_size, num_workers, metric_type)
+        dataset_results_dict = evaluate_linear_classifiers(
+            feature_model,
+            remove_ddp_wrapper(linear_classifiers),
+            test_data_loader,
+            metric_type,
+            metrics_file_path,
+            training_num_classes,
+            iteration,
+            prefixstring="",
+            class_mapping=class_mapping,
+            best_classifier_on_val=best_classifier_on_val,
+        )
+        results_dict[f"{test_dataset_str}_accuracy"] = 100.0 * dataset_results_dict["best_classifier"]["accuracy"]
+    return results_dict
+
+
+def run_eval_linear(
+    model,
+    output_dir,
+    train_dataset_str,
+    val_dataset_str,
+    batch_size,
+    epochs,
+    epoch_length,
+    num_workers,
+    save_checkpoint_frequency,
+    eval_period_iterations,
+    learning_rates,
+    autocast_dtype,
+    test_dataset_strs=None,
+    resume=True,
+    classifier_fpath=None,
+    val_class_mapping_fpath=None,
+    test_class_mapping_fpaths=[None],
+    val_metric_type=MetricType.MEAN_ACCURACY,
+    test_metric_types=None,
+):
+    seed = 0
+
+    if test_dataset_strs is None:
+        test_dataset_strs = [val_dataset_str]
+    if test_metric_types is None:
+        test_metric_types = [val_metric_type] * len(test_dataset_strs)
+    else:
+        assert len(test_metric_types) == len(test_dataset_strs)
+    assert len(test_dataset_strs) == len(test_class_mapping_fpaths)
+
+    train_transform = make_classification_train_transform()
+    train_dataset = make_dataset(
+        dataset_str=train_dataset_str,
+        transform=train_transform,
+    )
+    training_num_classes = len(torch.unique(torch.Tensor(train_dataset.get_targets().astype(int))))
+    sampler_type = SamplerType.SHARDED_INFINITE
+    # sampler_type = SamplerType.INFINITE
+
+    n_last_blocks_list = [1, 4]
+    n_last_blocks = max(n_last_blocks_list)
+    autocast_ctx = partial(torch.cuda.amp.autocast, enabled=True, dtype=autocast_dtype)
+    feature_model = ModelWithIntermediateLayers(model, n_last_blocks, autocast_ctx)
+    sample_output = feature_model(train_dataset[0][0].unsqueeze(0).cuda())
+
+    linear_classifiers, optim_param_groups = setup_linear_classifiers(
+        sample_output,
+        n_last_blocks_list,
+        learning_rates,
+        batch_size,
+        training_num_classes,
+    )
+
+    optimizer = torch.optim.SGD(optim_param_groups, momentum=0.9, weight_decay=0)
+    max_iter = epochs * epoch_length
+    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, max_iter, eta_min=0)
+    checkpointer = Checkpointer(linear_classifiers, output_dir, optimizer=optimizer, scheduler=scheduler)
+    start_iter = checkpointer.resume_or_load(classifier_fpath or "", resume=resume).get("iteration", -1) + 1
+    train_data_loader = make_data_loader(
+        dataset=train_dataset,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        shuffle=True,
+        seed=seed,
+        sampler_type=sampler_type,
+        sampler_advance=start_iter,
+        drop_last=True,
+        persistent_workers=True,
+    )
+    val_data_loader = make_eval_data_loader(val_dataset_str, batch_size, num_workers, val_metric_type)
+
+    checkpoint_period = save_checkpoint_frequency * epoch_length
+
+    if val_class_mapping_fpath is not None:
+        logger.info(f"Using class mapping from {val_class_mapping_fpath}")
+        val_class_mapping = np.load(val_class_mapping_fpath)
+    else:
+        val_class_mapping = None
+
+    test_class_mappings = []
+    for class_mapping_fpath in test_class_mapping_fpaths:
+        if class_mapping_fpath is not None and class_mapping_fpath != "None":
+            logger.info(f"Using class mapping from {class_mapping_fpath}")
+            class_mapping = np.load(class_mapping_fpath)
+        else:
+            class_mapping = None
+        test_class_mappings.append(class_mapping)
+
+    metrics_file_path = os.path.join(output_dir, "results_eval_linear.json")
+    val_results_dict, feature_model, linear_classifiers, iteration = eval_linear(
+        feature_model=feature_model,
+        linear_classifiers=linear_classifiers,
+        train_data_loader=train_data_loader,
+        val_data_loader=val_data_loader,
+        metrics_file_path=metrics_file_path,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        output_dir=output_dir,
+        max_iter=max_iter,
+        checkpoint_period=checkpoint_period,
+        running_checkpoint_period=epoch_length,
+        eval_period=eval_period_iterations,
+        metric_type=val_metric_type,
+        training_num_classes=training_num_classes,
+        resume=resume,
+        val_class_mapping=val_class_mapping,
+        classifier_fpath=classifier_fpath,
+    )
+    results_dict = {}
+    if len(test_dataset_strs) > 1 or test_dataset_strs[0] != val_dataset_str:
+        results_dict = test_on_datasets(
+            feature_model,
+            linear_classifiers,
+            test_dataset_strs,
+            batch_size,
+            0,  # num_workers,
+            test_metric_types,
+            metrics_file_path,
+            training_num_classes,
+            iteration,
+            val_results_dict["best_classifier"]["name"],
+            prefixstring="",
+            test_class_mappings=test_class_mappings,
+        )
+    results_dict["best_classifier"] = val_results_dict["best_classifier"]["name"]
+    results_dict[f"{val_dataset_str}_accuracy"] = 100.0 * val_results_dict["best_classifier"]["accuracy"]
+    logger.info("Test Results Dict " + str(results_dict))
+
+    return results_dict
+
+
+def main(args):
+    model, autocast_dtype = setup_and_build_model(args)
+    run_eval_linear(
+        model=model,
+        output_dir=args.output_dir,
+        train_dataset_str=args.train_dataset_str,
+        val_dataset_str=args.val_dataset_str,
+        test_dataset_strs=args.test_dataset_strs,
+        batch_size=args.batch_size,
+        epochs=args.epochs,
+        epoch_length=args.epoch_length,
+        num_workers=args.num_workers,
+        save_checkpoint_frequency=args.save_checkpoint_frequency,
+        eval_period_iterations=args.eval_period_iterations,
+        learning_rates=args.learning_rates,
+        autocast_dtype=autocast_dtype,
+        resume=not args.no_resume,
+        classifier_fpath=args.classifier_fpath,
+        val_metric_type=args.val_metric_type,
+        test_metric_types=args.test_metric_types,
+        val_class_mapping_fpath=args.val_class_mapping_fpath,
+        test_class_mapping_fpaths=args.test_class_mapping_fpaths,
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    description = "DINOv2 linear evaluation"
+    args_parser = get_args_parser(description=description)
+    args = args_parser.parse_args()
+    sys.exit(main(args))

torchhub/facebookresearch_dinov2_main/dinov2/eval/log_regression.py

@@ -0,0 +1,445 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import gc
+import logging
+import sys
+import time
+from typing import List, Optional
+
+from cuml.linear_model import LogisticRegression
+import torch
+import torch.backends.cudnn as cudnn
+import torch.distributed
+from torch import nn
+from torch.utils.data import TensorDataset
+from torchmetrics import MetricTracker
+
+from dinov2.data import make_dataset
+from dinov2.data.transforms import make_classification_eval_transform
+from dinov2.distributed import get_global_rank, get_global_size
+from dinov2.eval.metrics import MetricType, build_metric
+from dinov2.eval.setup import get_args_parser as get_setup_args_parser
+from dinov2.eval.setup import setup_and_build_model
+from dinov2.eval.utils import evaluate, extract_features
+from dinov2.utils.dtype import as_torch_dtype
+
+
+logger = logging.getLogger("dinov2")
+
+DEFAULT_MAX_ITER = 1_000
+C_POWER_RANGE = torch.linspace(-6, 5, 45)
+_CPU_DEVICE = torch.device("cpu")
+
+
+def get_args_parser(
+    description: Optional[str] = None,
+    parents: Optional[List[argparse.ArgumentParser]] = None,
+    add_help: bool = True,
+):
+    parents = parents or []
+    setup_args_parser = get_setup_args_parser(parents=parents, add_help=False)
+    parents = [setup_args_parser]
+    parser = argparse.ArgumentParser(
+        description=description,
+        parents=parents,
+        add_help=add_help,
+    )
+    parser.add_argument(
+        "--train-dataset",
+        dest="train_dataset_str",
+        type=str,
+        help="Training dataset",
+    )
+    parser.add_argument(
+        "--val-dataset",
+        dest="val_dataset_str",
+        type=str,
+        help="Validation dataset",
+    )
+    parser.add_argument(
+        "--finetune-dataset-str",
+        dest="finetune_dataset_str",
+        type=str,
+        help="Fine-tuning dataset",
+    )
+    parser.add_argument(
+        "--finetune-on-val",
+        action="store_true",
+        help="If there is no finetune dataset, whether to choose the "
+        "hyperparameters on the val set instead of 10%% of the train dataset",
+    )
+    parser.add_argument(
+        "--metric-type",
+        type=MetricType,
+        choices=list(MetricType),
+        help="Metric type",
+    )
+    parser.add_argument(
+        "--train-features-device",
+        type=str,
+        help="Device to gather train features (cpu, cuda, cuda:0, etc.), default: %(default)s",
+    )
+    parser.add_argument(
+        "--train-dtype",
+        type=str,
+        help="Data type to convert the train features to (default: %(default)s)",
+    )
+    parser.add_argument(
+        "--max-train-iters",
+        type=int,
+        help="Maximum number of train iterations (default: %(default)s)",
+    )
+    parser.set_defaults(
+        train_dataset_str="ImageNet:split=TRAIN",
+        val_dataset_str="ImageNet:split=VAL",
+        finetune_dataset_str=None,
+        metric_type=MetricType.MEAN_ACCURACY,
+        train_features_device="cpu",
+        train_dtype="float64",
+        max_train_iters=DEFAULT_MAX_ITER,
+        finetune_on_val=False,
+    )
+    return parser
+
+
+class LogRegModule(nn.Module):
+    def __init__(
+        self,
+        C,
+        max_iter=DEFAULT_MAX_ITER,
+        dtype=torch.float64,
+        device=_CPU_DEVICE,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.device = device
+        self.estimator = LogisticRegression(
+            penalty="l2",
+            C=C,
+            max_iter=max_iter,
+            output_type="numpy",
+            tol=1e-12,
+            linesearch_max_iter=50,
+        )
+
+    def forward(self, samples, targets):
+        samples_device = samples.device
+        samples = samples.to(dtype=self.dtype, device=self.device)
+        if self.device == _CPU_DEVICE:
+            samples = samples.numpy()
+        probas = self.estimator.predict_proba(samples)
+        return {"preds": torch.from_numpy(probas).to(samples_device), "target": targets}
+
+    def fit(self, train_features, train_labels):
+        train_features = train_features.to(dtype=self.dtype, device=self.device)
+        train_labels = train_labels.to(dtype=self.dtype, device=self.device)
+        if self.device == _CPU_DEVICE:
+            # both cuML and sklearn only work with numpy arrays on CPU
+            train_features = train_features.numpy()
+            train_labels = train_labels.numpy()
+        self.estimator.fit(train_features, train_labels)
+
+
+def evaluate_model(*, logreg_model, logreg_metric, test_data_loader, device):
+    postprocessors = {"metrics": logreg_model}
+    metrics = {"metrics": logreg_metric}
+    return evaluate(nn.Identity(), test_data_loader, postprocessors, metrics, device)
+
+
+def train_for_C(*, C, max_iter, train_features, train_labels, dtype=torch.float64, device=_CPU_DEVICE):
+    logreg_model = LogRegModule(C, max_iter=max_iter, dtype=dtype, device=device)
+    logreg_model.fit(train_features, train_labels)
+    return logreg_model
+
+
+def train_and_evaluate(
+    *,
+    C,
+    max_iter,
+    train_features,
+    train_labels,
+    logreg_metric,
+    test_data_loader,
+    train_dtype=torch.float64,
+    train_features_device,
+    eval_device,
+):
+    logreg_model = train_for_C(
+        C=C,
+        max_iter=max_iter,
+        train_features=train_features,
+        train_labels=train_labels,
+        dtype=train_dtype,
+        device=train_features_device,
+    )
+    return evaluate_model(
+        logreg_model=logreg_model,
+        logreg_metric=logreg_metric,
+        test_data_loader=test_data_loader,
+        device=eval_device,
+    )
+
+
+def sweep_C_values(
+    *,
+    train_features,
+    train_labels,
+    test_data_loader,
+    metric_type,
+    num_classes,
+    train_dtype=torch.float64,
+    train_features_device=_CPU_DEVICE,
+    max_train_iters=DEFAULT_MAX_ITER,
+):
+    if metric_type == MetricType.PER_CLASS_ACCURACY:
+        # If we want to output per-class accuracy, we select the hyperparameters with mean per class
+        metric_type = MetricType.MEAN_PER_CLASS_ACCURACY
+    logreg_metric = build_metric(metric_type, num_classes=num_classes)
+    metric_tracker = MetricTracker(logreg_metric, maximize=True)
+    ALL_C = 10**C_POWER_RANGE
+    logreg_models = {}
+
+    train_features = train_features.to(dtype=train_dtype, device=train_features_device)
+    train_labels = train_labels.to(device=train_features_device)
+
+    for i in range(get_global_rank(), len(ALL_C), get_global_size()):
+        C = ALL_C[i].item()
+        logger.info(
+            f"Training for C = {C:.5f}, dtype={train_dtype}, "
+            f"features: {train_features.shape}, {train_features.dtype}, "
+            f"labels: {train_labels.shape}, {train_labels.dtype}"
+        )
+        logreg_models[C] = train_for_C(
+            C=C,
+            max_iter=max_train_iters,
+            train_features=train_features,
+            train_labels=train_labels,
+            dtype=train_dtype,
+            device=train_features_device,
+        )
+
+    gather_list = [None for _ in range(get_global_size())]
+    torch.distributed.all_gather_object(gather_list, logreg_models)
+
+    logreg_models_gathered = {}
+    for logreg_dict in gather_list:
+        logreg_models_gathered.update(logreg_dict)
+
+    for i in range(len(ALL_C)):
+        metric_tracker.increment()
+        C = ALL_C[i].item()
+        evals = evaluate_model(
+            logreg_model=logreg_models_gathered[C],
+            logreg_metric=metric_tracker,
+            test_data_loader=test_data_loader,
+            device=torch.cuda.current_device(),
+        )
+        logger.info(f"Trained for C = {C:.5f}, accuracies = {evals}")
+
+        best_stats, which_epoch = metric_tracker.best_metric(return_step=True)
+        best_stats_100 = {k: 100.0 * v for k, v in best_stats.items()}
+        if which_epoch["top-1"] == i:
+            best_C = C
+    logger.info(f"Sweep best {best_stats_100}, best C = {best_C:.6f}")
+
+    return best_stats, best_C
+
+
+def eval_log_regression(
+    *,
+    model,
+    train_dataset,
+    val_dataset,
+    finetune_dataset,
+    metric_type,
+    batch_size,
+    num_workers,
+    finetune_on_val=False,
+    train_dtype=torch.float64,
+    train_features_device=_CPU_DEVICE,
+    max_train_iters=DEFAULT_MAX_ITER,
+):
+    """
+    Implements the "standard" process for log regression evaluation:
+    The value of C is chosen by training on train_dataset and evaluating on
+    finetune_dataset. Then, the final model is trained on a concatenation of
+    train_dataset and finetune_dataset, and is evaluated on val_dataset.
+    If there is no finetune_dataset, the value of C is the one that yields
+    the best results on a random 10% subset of the train dataset
+    """
+
+    start = time.time()
+
+    train_features, train_labels = extract_features(
+        model, train_dataset, batch_size, num_workers, gather_on_cpu=(train_features_device == _CPU_DEVICE)
+    )
+    val_features, val_labels = extract_features(
+        model, val_dataset, batch_size, num_workers, gather_on_cpu=(train_features_device == _CPU_DEVICE)
+    )
+    val_data_loader = torch.utils.data.DataLoader(
+        TensorDataset(val_features, val_labels),
+        batch_size=batch_size,
+        drop_last=False,
+        num_workers=0,
+        persistent_workers=False,
+    )
+
+    if finetune_dataset is None and finetune_on_val:
+        logger.info("Choosing hyperparameters on the val dataset")
+        finetune_features, finetune_labels = val_features, val_labels
+    elif finetune_dataset is None and not finetune_on_val:
+        logger.info("Choosing hyperparameters on 10% of the train dataset")
+        torch.manual_seed(0)
+        indices = torch.randperm(len(train_features), device=train_features.device)
+        finetune_index = indices[: len(train_features) // 10]
+        train_index = indices[len(train_features) // 10 :]
+        finetune_features, finetune_labels = train_features[finetune_index], train_labels[finetune_index]
+        train_features, train_labels = train_features[train_index], train_labels[train_index]
+    else:
+        logger.info("Choosing hyperparameters on the finetune dataset")
+        finetune_features, finetune_labels = extract_features(
+            model, finetune_dataset, batch_size, num_workers, gather_on_cpu=(train_features_device == _CPU_DEVICE)
+        )
+    # release the model - free GPU memory
+    del model
+    gc.collect()
+    torch.cuda.empty_cache()
+    finetune_data_loader = torch.utils.data.DataLoader(
+        TensorDataset(finetune_features, finetune_labels),
+        batch_size=batch_size,
+        drop_last=False,
+    )
+
+    if len(train_labels.shape) > 1:
+        num_classes = train_labels.shape[1]
+    else:
+        num_classes = train_labels.max() + 1
+
+    logger.info("Using cuML for logistic regression")
+
+    best_stats, best_C = sweep_C_values(
+        train_features=train_features,
+        train_labels=train_labels,
+        test_data_loader=finetune_data_loader,
+        metric_type=metric_type,
+        num_classes=num_classes,
+        train_dtype=train_dtype,
+        train_features_device=train_features_device,
+        max_train_iters=max_train_iters,
+    )
+
+    if not finetune_on_val:
+        logger.info("Best parameter found, concatenating features")
+        train_features = torch.cat((train_features, finetune_features))
+        train_labels = torch.cat((train_labels, finetune_labels))
+
+    logger.info("Training final model")
+    logreg_metric = build_metric(metric_type, num_classes=num_classes)
+    evals = train_and_evaluate(
+        C=best_C,
+        max_iter=max_train_iters,
+        train_features=train_features,
+        train_labels=train_labels,
+        logreg_metric=logreg_metric.clone(),
+        test_data_loader=val_data_loader,
+        eval_device=torch.cuda.current_device(),
+        train_dtype=train_dtype,
+        train_features_device=train_features_device,
+    )
+
+    best_stats = evals[1]["metrics"]
+
+    best_stats["best_C"] = best_C
+
+    logger.info(f"Log regression evaluation done in {int(time.time() - start)}s")
+    return best_stats
+
+
+def eval_log_regression_with_model(
+    model,
+    train_dataset_str="ImageNet:split=TRAIN",
+    val_dataset_str="ImageNet:split=VAL",
+    finetune_dataset_str=None,
+    autocast_dtype=torch.float,
+    finetune_on_val=False,
+    metric_type=MetricType.MEAN_ACCURACY,
+    train_dtype=torch.float64,
+    train_features_device=_CPU_DEVICE,
+    max_train_iters=DEFAULT_MAX_ITER,
+):
+    cudnn.benchmark = True
+
+    transform = make_classification_eval_transform(resize_size=224)
+    target_transform = None
+
+    train_dataset = make_dataset(dataset_str=train_dataset_str, transform=transform, target_transform=target_transform)
+    val_dataset = make_dataset(dataset_str=val_dataset_str, transform=transform, target_transform=target_transform)
+    if finetune_dataset_str is not None:
+        finetune_dataset = make_dataset(
+            dataset_str=finetune_dataset_str, transform=transform, target_transform=target_transform
+        )
+    else:
+        finetune_dataset = None
+
+    with torch.cuda.amp.autocast(dtype=autocast_dtype):
+        results_dict_logreg = eval_log_regression(
+            model=model,
+            train_dataset=train_dataset,
+            val_dataset=val_dataset,
+            finetune_dataset=finetune_dataset,
+            metric_type=metric_type,
+            batch_size=256,
+            num_workers=0,  # 5,
+            finetune_on_val=finetune_on_val,
+            train_dtype=train_dtype,
+            train_features_device=train_features_device,
+            max_train_iters=max_train_iters,
+        )
+
+    results_dict = {
+        "top-1": results_dict_logreg["top-1"].cpu().numpy() * 100.0,
+        "top-5": results_dict_logreg.get("top-5", torch.tensor(0.0)).cpu().numpy() * 100.0,
+        "best_C": results_dict_logreg["best_C"],
+    }
+    logger.info(
+        "\n".join(
+            [
+                "Training of the supervised logistic regression on frozen features completed.\n"
+                "Top-1 test accuracy: {acc:.1f}".format(acc=results_dict["top-1"]),
+                "Top-5 test accuracy: {acc:.1f}".format(acc=results_dict["top-5"]),
+                "obtained for C = {c:.6f}".format(c=results_dict["best_C"]),
+            ]
+        )
+    )
+
+    torch.distributed.barrier()
+    return results_dict
+
+
+def main(args):
+    model, autocast_dtype = setup_and_build_model(args)
+    eval_log_regression_with_model(
+        model=model,
+        train_dataset_str=args.train_dataset_str,
+        val_dataset_str=args.val_dataset_str,
+        finetune_dataset_str=args.finetune_dataset_str,
+        autocast_dtype=autocast_dtype,
+        finetune_on_val=args.finetune_on_val,
+        metric_type=args.metric_type,
+        train_dtype=as_torch_dtype(args.train_dtype),
+        train_features_device=torch.device(args.train_features_device),
+        max_train_iters=args.max_train_iters,
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    description = "DINOv2 logistic regression evaluation"
+    args_parser = get_args_parser(description=description)
+    args = args_parser.parse_args()
+    sys.exit(main(args))

torchhub/facebookresearch_dinov2_main/dinov2/eval/metrics.py

@@ -0,0 +1,114 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from enum import Enum
+import logging
+from typing import Any, Dict, Optional
+
+import torch
+from torch import Tensor
+from torchmetrics import Metric, MetricCollection
+from torchmetrics.classification import MulticlassAccuracy
+from torchmetrics.utilities.data import dim_zero_cat, select_topk
+
+
+logger = logging.getLogger("dinov2")
+
+
+class MetricType(Enum):
+    MEAN_ACCURACY = "mean_accuracy"
+    MEAN_PER_CLASS_ACCURACY = "mean_per_class_accuracy"
+    PER_CLASS_ACCURACY = "per_class_accuracy"
+    IMAGENET_REAL_ACCURACY = "imagenet_real_accuracy"
+
+    @property
+    def accuracy_averaging(self):
+        return getattr(AccuracyAveraging, self.name, None)
+
+    def __str__(self):
+        return self.value
+
+
+class AccuracyAveraging(Enum):
+    MEAN_ACCURACY = "micro"
+    MEAN_PER_CLASS_ACCURACY = "macro"
+    PER_CLASS_ACCURACY = "none"
+
+    def __str__(self):
+        return self.value
+
+
+def build_metric(metric_type: MetricType, *, num_classes: int, ks: Optional[tuple] = None):
+    if metric_type.accuracy_averaging is not None:
+        return build_topk_accuracy_metric(
+            average_type=metric_type.accuracy_averaging,
+            num_classes=num_classes,
+            ks=(1, 5) if ks is None else ks,
+        )
+    elif metric_type == MetricType.IMAGENET_REAL_ACCURACY:
+        return build_topk_imagenet_real_accuracy_metric(
+            num_classes=num_classes,
+            ks=(1, 5) if ks is None else ks,
+        )
+
+    raise ValueError(f"Unknown metric type {metric_type}")
+
+
+def build_topk_accuracy_metric(average_type: AccuracyAveraging, num_classes: int, ks: tuple = (1, 5)):
+    metrics: Dict[str, Metric] = {
+        f"top-{k}": MulticlassAccuracy(top_k=k, num_classes=int(num_classes), average=average_type.value) for k in ks
+    }
+    return MetricCollection(metrics)
+
+
+def build_topk_imagenet_real_accuracy_metric(num_classes: int, ks: tuple = (1, 5)):
+    metrics: Dict[str, Metric] = {f"top-{k}": ImageNetReaLAccuracy(top_k=k, num_classes=int(num_classes)) for k in ks}
+    return MetricCollection(metrics)
+
+
+class ImageNetReaLAccuracy(Metric):
+    is_differentiable: bool = False
+    higher_is_better: Optional[bool] = None
+    full_state_update: bool = False
+
+    def __init__(
+        self,
+        num_classes: int,
+        top_k: int = 1,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.num_classes = num_classes
+        self.top_k = top_k
+        self.add_state("tp", [], dist_reduce_fx="cat")
+
+    def update(self, preds: Tensor, target: Tensor) -> None:  # type: ignore
+        # preds [B, D]
+        # target [B, A]
+        # preds_oh [B, D] with 0 and 1
+        # select top K highest probabilities, use one hot representation
+        preds_oh = select_topk(preds, self.top_k)
+        # target_oh [B, D + 1] with 0 and 1
+        target_oh = torch.zeros((preds_oh.shape[0], preds_oh.shape[1] + 1), device=target.device, dtype=torch.int32)
+        target = target.long()
+        # for undefined targets (-1) use a fake value `num_classes`
+        target[target == -1] = self.num_classes
+        # fill targets, use one hot representation
+        target_oh.scatter_(1, target, 1)
+        # target_oh [B, D] (remove the fake target at index `num_classes`)
+        target_oh = target_oh[:, :-1]
+        # tp [B] with 0 and 1
+        tp = (preds_oh * target_oh == 1).sum(dim=1)
+        # at least one match between prediction and target
+        tp.clip_(max=1)
+        # ignore instances where no targets are defined
+        mask = target_oh.sum(dim=1) > 0
+        tp = tp[mask]
+        self.tp.append(tp)  # type: ignore
+
+    def compute(self) -> Tensor:
+        tp = dim_zero_cat(self.tp)  # type: ignore
+        return tp.float().mean()

torchhub/facebookresearch_dinov2_main/dinov2/eval/setup.py

@@ -0,0 +1,76 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+from typing import Any, List, Optional, Tuple
+
+import torch
+import torch.backends.cudnn as cudnn
+
+from dinov2.models import build_model_from_cfg
+from dinov2.utils.config import setup
+import dinov2.utils.utils as dinov2_utils
+
+
+def get_args_parser(
+    description: Optional[str] = None,
+    parents: Optional[List[argparse.ArgumentParser]] = None,
+    add_help: bool = True,
+):
+    parser = argparse.ArgumentParser(
+        description=description,
+        parents=parents or [],
+        add_help=add_help,
+    )
+    parser.add_argument(
+        "--config-file",
+        type=str,
+        help="Model configuration file",
+    )
+    parser.add_argument(
+        "--pretrained-weights",
+        type=str,
+        help="Pretrained model weights",
+    )
+    parser.add_argument(
+        "--output-dir",
+        default="",
+        type=str,
+        help="Output directory to write results and logs",
+    )
+    parser.add_argument(
+        "--opts",
+        help="Extra configuration options",
+        default=[],
+        nargs="+",
+    )
+    return parser
+
+
+def get_autocast_dtype(config):
+    teacher_dtype_str = config.compute_precision.teacher.backbone.mixed_precision.param_dtype
+    if teacher_dtype_str == "fp16":
+        return torch.half
+    elif teacher_dtype_str == "bf16":
+        return torch.bfloat16
+    else:
+        return torch.float
+
+
+def build_model_for_eval(config, pretrained_weights):
+    model, _ = build_model_from_cfg(config, only_teacher=True)
+    dinov2_utils.load_pretrained_weights(model, pretrained_weights, "teacher")
+    model.eval()
+    model.cuda()
+    return model
+
+
+def setup_and_build_model(args) -> Tuple[Any, torch.dtype]:
+    cudnn.benchmark = True
+    config = setup(args)
+    model = build_model_for_eval(config, args.pretrained_weights)
+    autocast_dtype = get_autocast_dtype(config)
+    return model, autocast_dtype

torchhub/facebookresearch_dinov2_main/dinov2/eval/utils.py

@@ -0,0 +1,147 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+from typing import Dict, Optional
+
+import torch
+from torch import nn
+from torchmetrics import MetricCollection
+
+from dinov2.data import DatasetWithEnumeratedTargets, SamplerType, make_data_loader
+import dinov2.distributed as distributed
+from dinov2.logging import MetricLogger
+
+
+logger = logging.getLogger("dinov2")
+
+
+class ModelWithNormalize(torch.nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.model = model
+
+    def forward(self, samples):
+        return nn.functional.normalize(self.model(samples), dim=1, p=2)
+
+
+class ModelWithIntermediateLayers(nn.Module):
+    def __init__(self, feature_model, n_last_blocks, autocast_ctx):
+        super().__init__()
+        self.feature_model = feature_model
+        self.feature_model.eval()
+        self.n_last_blocks = n_last_blocks
+        self.autocast_ctx = autocast_ctx
+
+    def forward(self, images):
+        with torch.inference_mode():
+            with self.autocast_ctx():
+                features = self.feature_model.get_intermediate_layers(
+                    images, self.n_last_blocks, return_class_token=True
+                )
+        return features
+
+
+@torch.inference_mode()
+def evaluate(
+    model: nn.Module,
+    data_loader,
+    postprocessors: Dict[str, nn.Module],
+    metrics: Dict[str, MetricCollection],
+    device: torch.device,
+    criterion: Optional[nn.Module] = None,
+):
+    model.eval()
+    if criterion is not None:
+        criterion.eval()
+
+    for metric in metrics.values():
+        metric = metric.to(device)
+
+    metric_logger = MetricLogger(delimiter="  ")
+    header = "Test:"
+
+    for samples, targets, *_ in metric_logger.log_every(data_loader, 10, header):
+        outputs = model(samples.to(device))
+        targets = targets.to(device)
+
+        if criterion is not None:
+            loss = criterion(outputs, targets)
+            metric_logger.update(loss=loss.item())
+
+        for k, metric in metrics.items():
+            metric_inputs = postprocessors[k](outputs, targets)
+            metric.update(**metric_inputs)
+
+    metric_logger.synchronize_between_processes()
+    logger.info(f"Averaged stats: {metric_logger}")
+
+    stats = {k: metric.compute() for k, metric in metrics.items()}
+    metric_logger_stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
+    return metric_logger_stats, stats
+
+
+def all_gather_and_flatten(tensor_rank):
+    tensor_all_ranks = torch.empty(
+        distributed.get_global_size(),
+        *tensor_rank.shape,
+        dtype=tensor_rank.dtype,
+        device=tensor_rank.device,
+    )
+    tensor_list = list(tensor_all_ranks.unbind(0))
+    torch.distributed.all_gather(tensor_list, tensor_rank.contiguous())
+    return tensor_all_ranks.flatten(end_dim=1)
+
+
+def extract_features(model, dataset, batch_size, num_workers, gather_on_cpu=False):
+    dataset_with_enumerated_targets = DatasetWithEnumeratedTargets(dataset)
+    sample_count = len(dataset_with_enumerated_targets)
+    data_loader = make_data_loader(
+        dataset=dataset_with_enumerated_targets,
+        batch_size=batch_size,
+        num_workers=num_workers,
+        sampler_type=SamplerType.DISTRIBUTED,
+        drop_last=False,
+        shuffle=False,
+    )
+    return extract_features_with_dataloader(model, data_loader, sample_count, gather_on_cpu)
+
+
+@torch.inference_mode()
+def extract_features_with_dataloader(model, data_loader, sample_count, gather_on_cpu=False):
+    gather_device = torch.device("cpu") if gather_on_cpu else torch.device("cuda")
+    metric_logger = MetricLogger(delimiter="  ")
+    features, all_labels = None, None
+    for samples, (index, labels_rank) in metric_logger.log_every(data_loader, 10):
+        samples = samples.cuda(non_blocking=True)
+        labels_rank = labels_rank.cuda(non_blocking=True)
+        index = index.cuda(non_blocking=True)
+        features_rank = model(samples).float()
+
+        # init storage feature matrix
+        if features is None:
+            features = torch.zeros(sample_count, features_rank.shape[-1], device=gather_device)
+            labels_shape = list(labels_rank.shape)
+            labels_shape[0] = sample_count
+            all_labels = torch.full(labels_shape, fill_value=-1, device=gather_device)
+            logger.info(f"Storing features into tensor of shape {features.shape}")
+
+        # share indexes, features and labels between processes
+        index_all = all_gather_and_flatten(index).to(gather_device)
+        features_all_ranks = all_gather_and_flatten(features_rank).to(gather_device)
+        labels_all_ranks = all_gather_and_flatten(labels_rank).to(gather_device)
+
+        # update storage feature matrix
+        if len(index_all) > 0:
+            features.index_copy_(0, index_all, features_all_ranks)
+            all_labels.index_copy_(0, index_all, labels_all_ranks)
+
+    logger.info(f"Features shape: {tuple(features.shape)}")
+    logger.info(f"Labels shape: {tuple(all_labels.shape)}")
+
+    assert torch.all(all_labels > -1)
+
+    return features, all_labels

torchhub/facebookresearch_dinov2_main/dinov2/fsdp/__init__.py

@@ -0,0 +1,158 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+from typing import Any
+
+import torch
+import dinov2.distributed as distributed
+from functools import partial
+from fvcore.common.checkpoint import Checkpointer
+from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp import ShardingStrategy
+from torch.distributed.fsdp import MixedPrecision
+from torch.distributed.fsdp import StateDictType
+from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
+from torch.distributed.fsdp.wrap import ModuleWrapPolicy
+from torch.distributed.fsdp._runtime_utils import _reshard
+
+
+def get_fsdp_wrapper(model_cfg, modules_to_wrap=set()):
+    sharding_strategy_dict = {
+        "NO_SHARD": ShardingStrategy.NO_SHARD,
+        "SHARD_GRAD_OP": ShardingStrategy.SHARD_GRAD_OP,
+        "FULL_SHARD": ShardingStrategy.FULL_SHARD,
+    }
+
+    dtype_dict = {
+        "fp32": torch.float32,
+        "fp16": torch.float16,
+        "bf16": torch.bfloat16,
+    }
+
+    mixed_precision_config = MixedPrecision(
+        param_dtype=dtype_dict[model_cfg.mixed_precision.param_dtype],
+        reduce_dtype=dtype_dict[model_cfg.mixed_precision.reduce_dtype],
+        buffer_dtype=dtype_dict[model_cfg.mixed_precision.buffer_dtype],
+    )
+
+    sharding_strategy_config = sharding_strategy_dict[model_cfg.sharding_strategy]
+
+    local_rank = distributed.get_local_rank()
+
+    fsdp_wrapper = partial(
+        FSDP,
+        sharding_strategy=sharding_strategy_config,
+        mixed_precision=mixed_precision_config,
+        device_id=local_rank,
+        sync_module_states=True,
+        use_orig_params=True,
+        auto_wrap_policy=ModuleWrapPolicy(modules_to_wrap),
+    )
+    return fsdp_wrapper
+
+
+def is_fsdp(x):
+    return isinstance(x, FSDP)
+
+
+def is_sharded_fsdp(x):
+    return is_fsdp(x) and x.sharding_strategy is not ShardingStrategy.NO_SHARD
+
+
+def free_if_fsdp(x):
+    if is_sharded_fsdp(x):
+        handles = x._handles
+        true_list = [True for h in handles]
+        _reshard(x, handles, true_list)
+
+
+def get_fsdp_modules(x):
+    return FSDP.fsdp_modules(x)
+
+
+def reshard_fsdp_model(x):
+    for m in get_fsdp_modules(x):
+        free_if_fsdp(m)
+
+
+def rankstr():
+    return f"rank_{distributed.get_global_rank()}"
+
+
+class FSDPCheckpointer(Checkpointer):
+    def save(self, name: str, **kwargs: Any) -> None:
+        """
+        Dump model and checkpointables to a file.
+
+        Args:
+            name (str): name of the file.
+            kwargs (dict): extra arbitrary data to save.
+        """
+        if not self.save_dir or not self.save_to_disk:
+            return
+
+        data = {}
+        with FSDP.state_dict_type(self.model, StateDictType.LOCAL_STATE_DICT):
+            data["model"] = self.model.state_dict()
+
+        # data["model"] = self.model.state_dict()
+        for key, obj in self.checkpointables.items():
+            data[key] = obj.state_dict()
+        data.update(kwargs)
+
+        basename = f"{name}.{rankstr()}.pth"
+        save_file = os.path.join(self.save_dir, basename)
+        assert os.path.basename(save_file) == basename, basename
+        self.logger.info("Saving checkpoint to {}".format(save_file))
+        with self.path_manager.open(save_file, "wb") as f:
+            torch.save(data, f)
+        self.tag_last_checkpoint(basename)
+
+    def load(self, *args, **kwargs):
+        with FSDP.state_dict_type(self.model, StateDictType.LOCAL_STATE_DICT):
+            return super().load(*args, **kwargs)
+
+    def has_checkpoint(self) -> bool:
+        """
+        Returns:
+            bool: whether a checkpoint exists in the target directory.
+        """
+        save_file = os.path.join(self.save_dir, f"last_checkpoint.{rankstr()}")
+        return self.path_manager.exists(save_file)
+
+    def get_checkpoint_file(self) -> str:
+        """
+        Returns:
+            str: The latest checkpoint file in target directory.
+        """
+        save_file = os.path.join(self.save_dir, f"last_checkpoint.{rankstr()}")
+        try:
+            with self.path_manager.open(save_file, "r") as f:
+                last_saved = f.read().strip()
+        except IOError:
+            # if file doesn't exist, maybe because it has just been
+            # deleted by a separate process
+            return ""
+        # pyre-fixme[6]: For 2nd param expected `Union[PathLike[str], str]` but got
+        #  `Union[bytes, str]`.
+        return os.path.join(self.save_dir, last_saved)
+
+    def tag_last_checkpoint(self, last_filename_basename: str) -> None:
+        """
+        Tag the last checkpoint.
+
+        Args:
+            last_filename_basename (str): the basename of the last filename.
+        """
+        if distributed.is_enabled():
+            torch.distributed.barrier()
+        save_file = os.path.join(self.save_dir, f"last_checkpoint.{rankstr()}")
+        with self.path_manager.open(save_file, "w") as f:
+            f.write(last_filename_basename)  # pyre-ignore
+
+
+ShardedGradScaler = ShardedGradScaler

torchhub/facebookresearch_dinov2_main/dinov2/layers/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .dino_head import DINOHead
+from .mlp import Mlp
+from .patch_embed import PatchEmbed
+from .swiglu_ffn import SwiGLUFFN, SwiGLUFFNFused
+from .block import NestedTensorBlock
+from .attention import MemEffAttention