Initial commit for NEMO tools FastAPI Space

app/dinov2/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+__version__ = "0.0.1"

app/dinov2/configs/__init__.py

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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)

app/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

app/dinov2/configs/eval/vitb14_reg4_pretrain.yaml

@@ -0,0 +1,9 @@
+student:
+  arch: vit_base
+  patch_size: 14
+  num_register_tokens: 4
+  interpolate_antialias: true
+  interpolate_offset: 0.0
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

app/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

app/dinov2/configs/eval/vitg14_reg4_pretrain.yaml

@@ -0,0 +1,10 @@
+student:
+  arch: vit_giant2
+  patch_size: 14
+  ffn_layer: swiglufused
+  num_register_tokens: 4
+  interpolate_antialias: true
+  interpolate_offset: 0.0
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

app/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

app/dinov2/configs/eval/vitl14_reg4_pretrain.yaml

@@ -0,0 +1,9 @@
+student:
+  arch: vit_large
+  patch_size: 14
+  num_register_tokens: 4
+  interpolate_antialias: true
+  interpolate_offset: 0.0
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

app/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

app/dinov2/configs/eval/vits14_reg4_pretrain.yaml

@@ -0,0 +1,9 @@
+student:
+  arch: vit_small
+  patch_size: 14
+  num_register_tokens: 4
+  interpolate_antialias: true
+  interpolate_offset: 0.0
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

app/dinov2/configs/ssl_default_config.yaml

@@ -0,0 +1,118 @@
+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
+  num_register_tokens: 0
+  interpolate_antialias: false
+  interpolate_offset: 0.1
+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

app/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

app/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

app/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

app/dinov2/data/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/adapters.py

@@ -0,0 +1,28 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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)

app/dinov2/data/augmentations.py

@@ -0,0 +1,118 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/collate.py

@@ -0,0 +1,49 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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),
+    }

app/dinov2/data/datasets/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .image_net import ImageNet
+from .image_net_22k import ImageNet22k

app/dinov2/data/datasets/decoders.py

@@ -0,0 +1,31 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/datasets/extended.py

@@ -0,0 +1,38 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/datasets/image_net.py

@@ -0,0 +1,290 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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()

app/dinov2/data/datasets/image_net_22k.py

@@ -0,0 +1,302 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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)

app/dinov2/data/loaders.py

@@ -0,0 +1,222 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/masking.py

@@ -0,0 +1,86 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/samplers.py

@@ -0,0 +1,229 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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

app/dinov2/data/transforms.py

@@ -0,0 +1,91 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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)

app/dinov2/distributed/__init__.py

@@ -0,0 +1,270 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# 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()

app/dinov2/eval/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.

app/dinov2/eval/depth/__init__.py

@@ -0,0 +1,4 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.

app/dinov2/eval/depth/models/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .backbones import *  # noqa: F403
+from .builder import BACKBONES, DEPTHER, HEADS, LOSSES, build_backbone, build_depther, build_head, build_loss
+from .decode_heads import *  # noqa: F403
+from .depther import *  # noqa: F403
+from .losses import *  # noqa: F403

app/dinov2/eval/depth/models/backbones/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .vision_transformer import DinoVisionTransformer

app/dinov2/eval/depth/models/backbones/vision_transformer.py

@@ -0,0 +1,16 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from mmcv.runner import BaseModule
+
+from ..builder import BACKBONES
+
+
+@BACKBONES.register_module()
+class DinoVisionTransformer(BaseModule):
+    """Vision Transformer."""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__()

app/dinov2/eval/depth/models/builder.py

@@ -0,0 +1,49 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import warnings
+
+from mmcv.cnn import MODELS as MMCV_MODELS
+from mmcv.cnn.bricks.registry import ATTENTION as MMCV_ATTENTION
+from mmcv.utils import Registry
+
+MODELS = Registry("models", parent=MMCV_MODELS)
+ATTENTION = Registry("attention", parent=MMCV_ATTENTION)
+
+
+BACKBONES = MODELS
+NECKS = MODELS
+HEADS = MODELS
+LOSSES = MODELS
+DEPTHER = MODELS
+
+
+def build_backbone(cfg):
+    """Build backbone."""
+    return BACKBONES.build(cfg)
+
+
+def build_neck(cfg):
+    """Build neck."""
+    return NECKS.build(cfg)
+
+
+def build_head(cfg):
+    """Build head."""
+    return HEADS.build(cfg)
+
+
+def build_loss(cfg):
+    """Build loss."""
+    return LOSSES.build(cfg)
+
+
+def build_depther(cfg, train_cfg=None, test_cfg=None):
+    """Build depther."""
+    if train_cfg is not None or test_cfg is not None:
+        warnings.warn("train_cfg and test_cfg is deprecated, " "please specify them in model", UserWarning)
+    assert cfg.get("train_cfg") is None or train_cfg is None, "train_cfg specified in both outer field and model field "
+    assert cfg.get("test_cfg") is None or test_cfg is None, "test_cfg specified in both outer field and model field "
+    return DEPTHER.build(cfg, default_args=dict(train_cfg=train_cfg, test_cfg=test_cfg))

app/dinov2/eval/depth/models/decode_heads/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .dpt_head import DPTHead
+from .linear_head import BNHead

app/dinov2/eval/depth/models/decode_heads/decode_head.py

@@ -0,0 +1,225 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import copy
+from abc import ABCMeta, abstractmethod
+
+import mmcv
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.runner import BaseModule, auto_fp16, force_fp32
+
+from ...ops import resize
+from ..builder import build_loss
+
+
+class DepthBaseDecodeHead(BaseModule, metaclass=ABCMeta):
+    """Base class for BaseDecodeHead.
+
+    Args:
+        in_channels (List): Input channels.
+        channels (int): Channels after modules, before conv_depth.
+        conv_cfg (dict|None): Config of conv layers. Default: None.
+        act_cfg (dict): Config of activation layers.
+            Default: dict(type='ReLU')
+        loss_decode (dict): Config of decode loss.
+            Default: dict(type='SigLoss').
+        sampler (dict|None): The config of depth map sampler.
+            Default: None.
+        align_corners (bool): align_corners argument of F.interpolate.
+            Default: False.
+        min_depth (int): Min depth in dataset setting.
+            Default: 1e-3.
+        max_depth (int): Max depth in dataset setting.
+            Default: None.
+        norm_cfg (dict|None): Config of norm layers.
+            Default: None.
+        classify (bool): Whether predict depth in a cls.-reg. manner.
+            Default: False.
+        n_bins (int): The number of bins used in cls. step.
+            Default: 256.
+        bins_strategy (str): The discrete strategy used in cls. step.
+            Default: 'UD'.
+        norm_strategy (str): The norm strategy on cls. probability
+            distribution. Default: 'linear'
+        scale_up (str): Whether predict depth in a scale-up manner.
+            Default: False.
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        channels=96,
+        conv_cfg=None,
+        act_cfg=dict(type="ReLU"),
+        loss_decode=dict(type="SigLoss", valid_mask=True, loss_weight=10),
+        sampler=None,
+        align_corners=False,
+        min_depth=1e-3,
+        max_depth=None,
+        norm_cfg=None,
+        classify=False,
+        n_bins=256,
+        bins_strategy="UD",
+        norm_strategy="linear",
+        scale_up=False,
+    ):
+        super(DepthBaseDecodeHead, self).__init__()
+
+        self.in_channels = in_channels
+        self.channels = channels
+        self.conv_cfg = conv_cfg
+        self.act_cfg = act_cfg
+        if isinstance(loss_decode, dict):
+            self.loss_decode = build_loss(loss_decode)
+        elif isinstance(loss_decode, (list, tuple)):
+            self.loss_decode = nn.ModuleList()
+            for loss in loss_decode:
+                self.loss_decode.append(build_loss(loss))
+        self.align_corners = align_corners
+        self.min_depth = min_depth
+        self.max_depth = max_depth
+        self.norm_cfg = norm_cfg
+        self.classify = classify
+        self.n_bins = n_bins
+        self.scale_up = scale_up
+
+        if self.classify:
+            assert bins_strategy in ["UD", "SID"], "Support bins_strategy: UD, SID"
+            assert norm_strategy in ["linear", "softmax", "sigmoid"], "Support norm_strategy: linear, softmax, sigmoid"
+
+            self.bins_strategy = bins_strategy
+            self.norm_strategy = norm_strategy
+            self.softmax = nn.Softmax(dim=1)
+            self.conv_depth = nn.Conv2d(channels, n_bins, kernel_size=3, padding=1, stride=1)
+        else:
+            self.conv_depth = nn.Conv2d(channels, 1, kernel_size=3, padding=1, stride=1)
+
+        self.fp16_enabled = False
+        self.relu = nn.ReLU()
+        self.sigmoid = nn.Sigmoid()
+
+    def extra_repr(self):
+        """Extra repr."""
+        s = f"align_corners={self.align_corners}"
+        return s
+
+    @auto_fp16()
+    @abstractmethod
+    def forward(self, inputs, img_metas):
+        """Placeholder of forward function."""
+        pass
+
+    def forward_train(self, img, inputs, img_metas, depth_gt, train_cfg):
+        """Forward function for training.
+        Args:
+            inputs (list[Tensor]): List of multi-level img features.
+            img_metas (list[dict]): List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `depth/datasets/pipelines/formatting.py:Collect`.
+            depth_gt (Tensor): GT depth
+            train_cfg (dict): The training config.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+        depth_pred = self.forward(inputs, img_metas)
+        losses = self.losses(depth_pred, depth_gt)
+
+        log_imgs = self.log_images(img[0], depth_pred[0], depth_gt[0], img_metas[0])
+        losses.update(**log_imgs)
+
+        return losses
+
+    def forward_test(self, inputs, img_metas, test_cfg):
+        """Forward function for testing.
+        Args:
+            inputs (list[Tensor]): List of multi-level img features.
+            img_metas (list[dict]): List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `depth/datasets/pipelines/formatting.py:Collect`.
+            test_cfg (dict): The testing config.
+
+        Returns:
+            Tensor: Output depth map.
+        """
+        return self.forward(inputs, img_metas)
+
+    def depth_pred(self, feat):
+        """Prediction each pixel."""
+        if self.classify:
+            logit = self.conv_depth(feat)
+
+            if self.bins_strategy == "UD":
+                bins = torch.linspace(self.min_depth, self.max_depth, self.n_bins, device=feat.device)
+            elif self.bins_strategy == "SID":
+                bins = torch.logspace(self.min_depth, self.max_depth, self.n_bins, device=feat.device)
+
+            # following Adabins, default linear
+            if self.norm_strategy == "linear":
+                logit = torch.relu(logit)
+                eps = 0.1
+                logit = logit + eps
+                logit = logit / logit.sum(dim=1, keepdim=True)
+            elif self.norm_strategy == "softmax":
+                logit = torch.softmax(logit, dim=1)
+            elif self.norm_strategy == "sigmoid":
+                logit = torch.sigmoid(logit)
+                logit = logit / logit.sum(dim=1, keepdim=True)
+
+            output = torch.einsum("ikmn,k->imn", [logit, bins]).unsqueeze(dim=1)
+
+        else:
+            if self.scale_up:
+                output = self.sigmoid(self.conv_depth(feat)) * self.max_depth
+            else:
+                output = self.relu(self.conv_depth(feat)) + self.min_depth
+        return output
+
+    @force_fp32(apply_to=("depth_pred",))
+    def losses(self, depth_pred, depth_gt):
+        """Compute depth loss."""
+        loss = dict()
+        depth_pred = resize(
+            input=depth_pred, size=depth_gt.shape[2:], mode="bilinear", align_corners=self.align_corners, warning=False
+        )
+        if not isinstance(self.loss_decode, nn.ModuleList):
+            losses_decode = [self.loss_decode]
+        else:
+            losses_decode = self.loss_decode
+        for loss_decode in losses_decode:
+            if loss_decode.loss_name not in loss:
+                loss[loss_decode.loss_name] = loss_decode(depth_pred, depth_gt)
+            else:
+                loss[loss_decode.loss_name] += loss_decode(depth_pred, depth_gt)
+        return loss
+
+    def log_images(self, img_path, depth_pred, depth_gt, img_meta):
+        show_img = copy.deepcopy(img_path.detach().cpu().permute(1, 2, 0))
+        show_img = show_img.numpy().astype(np.float32)
+        show_img = mmcv.imdenormalize(
+            show_img,
+            img_meta["img_norm_cfg"]["mean"],
+            img_meta["img_norm_cfg"]["std"],
+            img_meta["img_norm_cfg"]["to_rgb"],
+        )
+        show_img = np.clip(show_img, 0, 255)
+        show_img = show_img.astype(np.uint8)
+        show_img = show_img[:, :, ::-1]
+        show_img = show_img.transpose(0, 2, 1)
+        show_img = show_img.transpose(1, 0, 2)
+
+        depth_pred = depth_pred / torch.max(depth_pred)
+        depth_gt = depth_gt / torch.max(depth_gt)
+
+        depth_pred_color = copy.deepcopy(depth_pred.detach().cpu())
+        depth_gt_color = copy.deepcopy(depth_gt.detach().cpu())
+
+        return {"img_rgb": show_img, "img_depth_pred": depth_pred_color, "img_depth_gt": depth_gt_color}

app/dinov2/eval/depth/models/decode_heads/dpt_head.py

@@ -0,0 +1,270 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule, Linear, build_activation_layer
+from mmcv.runner import BaseModule
+
+from ...ops import resize
+from ..builder import HEADS
+from .decode_head import DepthBaseDecodeHead
+
+
+class Interpolate(nn.Module):
+    def __init__(self, scale_factor, mode, align_corners=False):
+        super(Interpolate, self).__init__()
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        x = self.interp(x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners)
+        return x
+
+
+class HeadDepth(nn.Module):
+    def __init__(self, features):
+        super(HeadDepth, self).__init__()
+        self.head = nn.Sequential(
+            nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+        )
+
+    def forward(self, x):
+        x = self.head(x)
+        return x
+
+
+class ReassembleBlocks(BaseModule):
+    """ViTPostProcessBlock, process cls_token in ViT backbone output and
+    rearrange the feature vector to feature map.
+    Args:
+        in_channels (int): ViT feature channels. Default: 768.
+        out_channels (List): output channels of each stage.
+            Default: [96, 192, 384, 768].
+        readout_type (str): Type of readout operation. Default: 'ignore'.
+        patch_size (int): The patch size. Default: 16.
+        init_cfg (dict, optional): Initialization config dict. Default: None.
+    """
+
+    def __init__(
+        self, in_channels=768, out_channels=[96, 192, 384, 768], readout_type="ignore", patch_size=16, init_cfg=None
+    ):
+        super(ReassembleBlocks, self).__init__(init_cfg)
+
+        assert readout_type in ["ignore", "add", "project"]
+        self.readout_type = readout_type
+        self.patch_size = patch_size
+
+        self.projects = nn.ModuleList(
+            [
+                ConvModule(
+                    in_channels=in_channels,
+                    out_channels=out_channel,
+                    kernel_size=1,
+                    act_cfg=None,
+                )
+                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 self.readout_type == "project":
+            self.readout_projects = nn.ModuleList()
+            for _ in range(len(self.projects)):
+                self.readout_projects.append(
+                    nn.Sequential(Linear(2 * in_channels, in_channels), build_activation_layer(dict(type="GELU")))
+                )
+
+    def forward(self, inputs):
+        assert isinstance(inputs, list)
+        out = []
+        for i, x in enumerate(inputs):
+            assert len(x) == 2
+            x, cls_token = x[0], x[1]
+            feature_shape = x.shape
+            if self.readout_type == "project":
+                x = x.flatten(2).permute((0, 2, 1))
+                readout = cls_token.unsqueeze(1).expand_as(x)
+                x = self.readout_projects[i](torch.cat((x, readout), -1))
+                x = x.permute(0, 2, 1).reshape(feature_shape)
+            elif self.readout_type == "add":
+                x = x.flatten(2) + cls_token.unsqueeze(-1)
+                x = x.reshape(feature_shape)
+            else:
+                pass
+            x = self.projects[i](x)
+            x = self.resize_layers[i](x)
+            out.append(x)
+        return out
+
+
+class PreActResidualConvUnit(BaseModule):
+    """ResidualConvUnit, pre-activate residual unit.
+    Args:
+        in_channels (int): number of channels in the input feature map.
+        act_cfg (dict): dictionary to construct and config activation layer.
+        norm_cfg (dict): dictionary to construct and config norm layer.
+        stride (int): stride of the first block. Default: 1
+        dilation (int): dilation rate for convs layers. Default: 1.
+        init_cfg (dict, optional): Initialization config dict. Default: None.
+    """
+
+    def __init__(self, in_channels, act_cfg, norm_cfg, stride=1, dilation=1, init_cfg=None):
+        super(PreActResidualConvUnit, self).__init__(init_cfg)
+
+        self.conv1 = ConvModule(
+            in_channels,
+            in_channels,
+            3,
+            stride=stride,
+            padding=dilation,
+            dilation=dilation,
+            norm_cfg=norm_cfg,
+            act_cfg=act_cfg,
+            bias=False,
+            order=("act", "conv", "norm"),
+        )
+
+        self.conv2 = ConvModule(
+            in_channels,
+            in_channels,
+            3,
+            padding=1,
+            norm_cfg=norm_cfg,
+            act_cfg=act_cfg,
+            bias=False,
+            order=("act", "conv", "norm"),
+        )
+
+    def forward(self, inputs):
+        inputs_ = inputs.clone()
+        x = self.conv1(inputs)
+        x = self.conv2(x)
+        return x + inputs_
+
+
+class FeatureFusionBlock(BaseModule):
+    """FeatureFusionBlock, merge feature map from different stages.
+    Args:
+        in_channels (int): Input channels.
+        act_cfg (dict): The activation config for ResidualConvUnit.
+        norm_cfg (dict): Config dict for normalization layer.
+        expand (bool): Whether expand the channels in post process block.
+            Default: False.
+        align_corners (bool): align_corner setting for bilinear upsample.
+            Default: True.
+        init_cfg (dict, optional): Initialization config dict. Default: None.
+    """
+
+    def __init__(self, in_channels, act_cfg, norm_cfg, expand=False, align_corners=True, init_cfg=None):
+        super(FeatureFusionBlock, self).__init__(init_cfg)
+
+        self.in_channels = in_channels
+        self.expand = expand
+        self.align_corners = align_corners
+
+        self.out_channels = in_channels
+        if self.expand:
+            self.out_channels = in_channels // 2
+
+        self.project = ConvModule(self.in_channels, self.out_channels, kernel_size=1, act_cfg=None, bias=True)
+
+        self.res_conv_unit1 = PreActResidualConvUnit(in_channels=self.in_channels, act_cfg=act_cfg, norm_cfg=norm_cfg)
+        self.res_conv_unit2 = PreActResidualConvUnit(in_channels=self.in_channels, act_cfg=act_cfg, norm_cfg=norm_cfg)
+
+    def forward(self, *inputs):
+        x = inputs[0]
+        if len(inputs) == 2:
+            if x.shape != inputs[1].shape:
+                res = resize(inputs[1], size=(x.shape[2], x.shape[3]), mode="bilinear", align_corners=False)
+            else:
+                res = inputs[1]
+            x = x + self.res_conv_unit1(res)
+        x = self.res_conv_unit2(x)
+        x = resize(x, scale_factor=2, mode="bilinear", align_corners=self.align_corners)
+        x = self.project(x)
+        return x
+
+
+@HEADS.register_module()
+class DPTHead(DepthBaseDecodeHead):
+    """Vision Transformers for Dense Prediction.
+    This head is implemented of `DPT <https://arxiv.org/abs/2103.13413>`_.
+    Args:
+        embed_dims (int): The embed dimension of the ViT backbone.
+            Default: 768.
+        post_process_channels (List): Out channels of post process conv
+            layers. Default: [96, 192, 384, 768].
+        readout_type (str): Type of readout operation. Default: 'ignore'.
+        patch_size (int): The patch size. Default: 16.
+        expand_channels (bool): Whether expand the channels in post process
+            block. Default: False.
+    """
+
+    def __init__(
+        self,
+        embed_dims=768,
+        post_process_channels=[96, 192, 384, 768],
+        readout_type="ignore",
+        patch_size=16,
+        expand_channels=False,
+        **kwargs
+    ):
+        super(DPTHead, self).__init__(**kwargs)
+
+        self.in_channels = self.in_channels
+        self.expand_channels = expand_channels
+        self.reassemble_blocks = ReassembleBlocks(embed_dims, post_process_channels, readout_type, patch_size)
+
+        self.post_process_channels = [
+            channel * math.pow(2, i) if expand_channels else channel for i, channel in enumerate(post_process_channels)
+        ]
+        self.convs = nn.ModuleList()
+        for channel in self.post_process_channels:
+            self.convs.append(ConvModule(channel, self.channels, kernel_size=3, padding=1, act_cfg=None, bias=False))
+        self.fusion_blocks = nn.ModuleList()
+        for _ in range(len(self.convs)):
+            self.fusion_blocks.append(FeatureFusionBlock(self.channels, self.act_cfg, self.norm_cfg))
+        self.fusion_blocks[0].res_conv_unit1 = None
+        self.project = ConvModule(self.channels, self.channels, kernel_size=3, padding=1, norm_cfg=self.norm_cfg)
+        self.num_fusion_blocks = len(self.fusion_blocks)
+        self.num_reassemble_blocks = len(self.reassemble_blocks.resize_layers)
+        self.num_post_process_channels = len(self.post_process_channels)
+        assert self.num_fusion_blocks == self.num_reassemble_blocks
+        assert self.num_reassemble_blocks == self.num_post_process_channels
+        self.conv_depth = HeadDepth(self.channels)
+
+    def forward(self, inputs, img_metas):
+        assert len(inputs) == self.num_reassemble_blocks
+        x = [inp for inp in inputs]
+        x = self.reassemble_blocks(x)
+        x = [self.convs[i](feature) for i, feature in enumerate(x)]
+        out = self.fusion_blocks[0](x[-1])
+        for i in range(1, len(self.fusion_blocks)):
+            out = self.fusion_blocks[i](out, x[-(i + 1)])
+        out = self.project(out)
+        out = self.depth_pred(out)
+        return out

app/dinov2/eval/depth/models/decode_heads/linear_head.py

@@ -0,0 +1,89 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+
+from ...ops import resize
+from ..builder import HEADS
+from .decode_head import DepthBaseDecodeHead
+
+
+@HEADS.register_module()
+class BNHead(DepthBaseDecodeHead):
+    """Just a batchnorm."""
+
+    def __init__(self, input_transform="resize_concat", in_index=(0, 1, 2, 3), upsample=1, **kwargs):
+        super().__init__(**kwargs)
+        self.input_transform = input_transform
+        self.in_index = in_index
+        self.upsample = upsample
+        # self.bn = nn.SyncBatchNorm(self.in_channels)
+        if self.classify:
+            self.conv_depth = nn.Conv2d(self.channels, self.n_bins, kernel_size=1, padding=0, stride=1)
+        else:
+            self.conv_depth = nn.Conv2d(self.channels, 1, kernel_size=1, padding=0, stride=1)
+
+    def _transform_inputs(self, inputs):
+        """Transform inputs for decoder.
+        Args:
+            inputs (list[Tensor]): List of multi-level img features.
+        Returns:
+            Tensor: The transformed inputs
+        """
+
+        if "concat" in self.input_transform:
+            inputs = [inputs[i] for i in self.in_index]
+            if "resize" in self.input_transform:
+                inputs = [
+                    resize(
+                        input=x,
+                        size=[s * self.upsample for s in inputs[0].shape[2:]],
+                        mode="bilinear",
+                        align_corners=self.align_corners,
+                    )
+                    for x in inputs
+                ]
+            inputs = torch.cat(inputs, dim=1)
+        elif self.input_transform == "multiple_select":
+            inputs = [inputs[i] for i in self.in_index]
+        else:
+            inputs = inputs[self.in_index]
+
+        return inputs
+
+    def _forward_feature(self, inputs, img_metas=None, **kwargs):
+        """Forward function for feature maps before classifying each pixel with
+        ``self.cls_seg`` fc.
+        Args:
+            inputs (list[Tensor]): List of multi-level img features.
+        Returns:
+            feats (Tensor): A tensor of shape (batch_size, self.channels,
+                H, W) which is feature map for last layer of decoder head.
+        """
+        # accept lists (for cls token)
+        inputs = list(inputs)
+        for i, x in enumerate(inputs):
+            if len(x) == 2:
+                x, cls_token = x[0], x[1]
+                if len(x.shape) == 2:
+                    x = x[:, :, None, None]
+                cls_token = cls_token[:, :, None, None].expand_as(x)
+                inputs[i] = torch.cat((x, cls_token), 1)
+            else:
+                x = x[0]
+                if len(x.shape) == 2:
+                    x = x[:, :, None, None]
+                inputs[i] = x
+        x = self._transform_inputs(inputs)
+        # feats = self.bn(x)
+        return x
+
+    def forward(self, inputs, img_metas=None, **kwargs):
+        """Forward function."""
+        output = self._forward_feature(inputs, img_metas=img_metas, **kwargs)
+        output = self.depth_pred(output)
+
+        return output

app/dinov2/eval/depth/models/depther/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .base import BaseDepther
+from .encoder_decoder import DepthEncoderDecoder

app/dinov2/eval/depth/models/depther/base.py

@@ -0,0 +1,194 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from abc import ABCMeta, abstractmethod
+from collections import OrderedDict
+
+import torch
+import torch.distributed as dist
+from mmcv.runner import BaseModule, auto_fp16
+
+
+class BaseDepther(BaseModule, metaclass=ABCMeta):
+    """Base class for depther."""
+
+    def __init__(self, init_cfg=None):
+        super(BaseDepther, self).__init__(init_cfg)
+        self.fp16_enabled = False
+
+    @property
+    def with_neck(self):
+        """bool: whether the depther has neck"""
+        return hasattr(self, "neck") and self.neck is not None
+
+    @property
+    def with_auxiliary_head(self):
+        """bool: whether the depther has auxiliary head"""
+        return hasattr(self, "auxiliary_head") and self.auxiliary_head is not None
+
+    @property
+    def with_decode_head(self):
+        """bool: whether the depther has decode head"""
+        return hasattr(self, "decode_head") and self.decode_head is not None
+
+    @abstractmethod
+    def extract_feat(self, imgs):
+        """Placeholder for extract features from images."""
+        pass
+
+    @abstractmethod
+    def encode_decode(self, img, img_metas):
+        """Placeholder for encode images with backbone and decode into a
+        semantic depth map of the same size as input."""
+        pass
+
+    @abstractmethod
+    def forward_train(self, imgs, img_metas, **kwargs):
+        """Placeholder for Forward function for training."""
+        pass
+
+    @abstractmethod
+    def simple_test(self, img, img_meta, **kwargs):
+        """Placeholder for single image test."""
+        pass
+
+    @abstractmethod
+    def aug_test(self, imgs, img_metas, **kwargs):
+        """Placeholder for augmentation test."""
+        pass
+
+    def forward_test(self, imgs, img_metas, **kwargs):
+        """
+        Args:
+            imgs (List[Tensor]): the outer list indicates test-time
+                augmentations and inner Tensor should have a shape NxCxHxW,
+                which contains all images in the batch.
+            img_metas (List[List[dict]]): the outer list indicates test-time
+                augs (multiscale, flip, etc.) and the inner list indicates
+                images in a batch.
+        """
+        for var, name in [(imgs, "imgs"), (img_metas, "img_metas")]:
+            if not isinstance(var, list):
+                raise TypeError(f"{name} must be a list, but got " f"{type(var)}")
+        num_augs = len(imgs)
+        if num_augs != len(img_metas):
+            raise ValueError(f"num of augmentations ({len(imgs)}) != " f"num of image meta ({len(img_metas)})")
+        # all images in the same aug batch all of the same ori_shape and pad
+        # shape
+        for img_meta in img_metas:
+            ori_shapes = [_["ori_shape"] for _ in img_meta]
+            assert all(shape == ori_shapes[0] for shape in ori_shapes)
+            img_shapes = [_["img_shape"] for _ in img_meta]
+            assert all(shape == img_shapes[0] for shape in img_shapes)
+            pad_shapes = [_["pad_shape"] for _ in img_meta]
+            assert all(shape == pad_shapes[0] for shape in pad_shapes)
+
+        if num_augs == 1:
+            return self.simple_test(imgs[0], img_metas[0], **kwargs)
+        else:
+            return self.aug_test(imgs, img_metas, **kwargs)
+
+    @auto_fp16(apply_to=("img",))
+    def forward(self, img, img_metas, return_loss=True, **kwargs):
+        """Calls either :func:`forward_train` or :func:`forward_test` depending
+        on whether ``return_loss`` is ``True``.
+
+        Note this setting will change the expected inputs. When
+        ``return_loss=True``, img and img_meta are single-nested (i.e. Tensor
+        and List[dict]), and when ``resturn_loss=False``, img and img_meta
+        should be double nested (i.e.  List[Tensor], List[List[dict]]), with
+        the outer list indicating test time augmentations.
+        """
+        if return_loss:
+            return self.forward_train(img, img_metas, **kwargs)
+        else:
+            return self.forward_test(img, img_metas, **kwargs)
+
+    def train_step(self, data_batch, optimizer, **kwargs):
+        """The iteration step during training.
+
+        This method defines an iteration step during training, except for the
+        back propagation and optimizer updating, which are done in an optimizer
+        hook. Note that in some complicated cases or models, the whole process
+        including back propagation and optimizer updating is also defined in
+        this method, such as GAN.
+
+        Args:
+            data (dict): The output of dataloader.
+            optimizer (:obj:`torch.optim.Optimizer` | dict): The optimizer of
+                runner is passed to ``train_step()``. This argument is unused
+                and reserved.
+
+        Returns:
+            dict: It should contain at least 3 keys: ``loss``, ``log_vars``,
+                ``num_samples``.
+                ``loss`` is a tensor for back propagation, which can be a
+                weighted sum of multiple losses.
+                ``log_vars`` contains all the variables to be sent to the
+                logger.
+                ``num_samples`` indicates the batch size (when the model is
+                DDP, it means the batch size on each GPU), which is used for
+                averaging the logs.
+        """
+        losses = self(**data_batch)
+
+        # split losses and images
+        real_losses = {}
+        log_imgs = {}
+        for k, v in losses.items():
+            if "img" in k:
+                log_imgs[k] = v
+            else:
+                real_losses[k] = v
+
+        loss, log_vars = self._parse_losses(real_losses)
+
+        outputs = dict(loss=loss, log_vars=log_vars, num_samples=len(data_batch["img_metas"]), log_imgs=log_imgs)
+
+        return outputs
+
+    def val_step(self, data_batch, **kwargs):
+        """The iteration step during validation.
+
+        This method shares the same signature as :func:`train_step`, but used
+        during val epochs. Note that the evaluation after training epochs is
+        not implemented with this method, but an evaluation hook.
+        """
+        output = self(**data_batch, **kwargs)
+        return output
+
+    @staticmethod
+    def _parse_losses(losses):
+        """Parse the raw outputs (losses) of the network.
+
+        Args:
+            losses (dict): Raw output of the network, which usually contain
+                losses and other necessary information.
+
+        Returns:
+            tuple[Tensor, dict]: (loss, log_vars), loss is the loss tensor
+                which may be a weighted sum of all losses, log_vars contains
+                all the variables to be sent to the logger.
+        """
+        log_vars = OrderedDict()
+        for loss_name, loss_value in losses.items():
+            if isinstance(loss_value, torch.Tensor):
+                log_vars[loss_name] = loss_value.mean()
+            elif isinstance(loss_value, list):
+                log_vars[loss_name] = sum(_loss.mean() for _loss in loss_value)
+            else:
+                raise TypeError(f"{loss_name} is not a tensor or list of tensors")
+
+        loss = sum(_value for _key, _value in log_vars.items() if "loss" in _key)
+
+        log_vars["loss"] = loss
+        for loss_name, loss_value in log_vars.items():
+            # reduce loss when distributed training
+            if dist.is_available() and dist.is_initialized():
+                loss_value = loss_value.data.clone()
+                dist.all_reduce(loss_value.div_(dist.get_world_size()))
+            log_vars[loss_name] = loss_value.item()
+
+        return loss, log_vars

app/dinov2/eval/depth/models/depther/encoder_decoder.py

@@ -0,0 +1,236 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn.functional as F
+
+from ...models import builder
+from ...models.builder import DEPTHER
+from ...ops import resize
+from .base import BaseDepther
+
+
+def add_prefix(inputs, prefix):
+    """Add prefix for dict.
+
+    Args:
+        inputs (dict): The input dict with str keys.
+        prefix (str): The prefix to add.
+
+    Returns:
+
+        dict: The dict with keys updated with ``prefix``.
+    """
+
+    outputs = dict()
+    for name, value in inputs.items():
+        outputs[f"{prefix}.{name}"] = value
+
+    return outputs
+
+
+@DEPTHER.register_module()
+class DepthEncoderDecoder(BaseDepther):
+    """Encoder Decoder depther.
+
+    EncoderDecoder typically consists of backbone, (neck) and decode_head.
+    """
+
+    def __init__(self, backbone, decode_head, neck=None, train_cfg=None, test_cfg=None, pretrained=None, init_cfg=None):
+        super(DepthEncoderDecoder, self).__init__(init_cfg)
+        if pretrained is not None:
+            assert backbone.get("pretrained") is None, "both backbone and depther set pretrained weight"
+            backbone.pretrained = pretrained
+        self.backbone = builder.build_backbone(backbone)
+        self._init_decode_head(decode_head)
+
+        if neck is not None:
+            self.neck = builder.build_neck(neck)
+
+        self.train_cfg = train_cfg
+        self.test_cfg = test_cfg
+
+        assert self.with_decode_head
+
+    def _init_decode_head(self, decode_head):
+        """Initialize ``decode_head``"""
+        self.decode_head = builder.build_head(decode_head)
+        self.align_corners = self.decode_head.align_corners
+
+    def extract_feat(self, img):
+        """Extract features from images."""
+        x = self.backbone(img)
+        if self.with_neck:
+            x = self.neck(x)
+        return x
+
+    def encode_decode(self, img, img_metas, rescale=True, size=None):
+        """Encode images with backbone and decode into a depth estimation
+        map of the same size as input."""
+        x = self.extract_feat(img)
+        out = self._decode_head_forward_test(x, img_metas)
+        # crop the pred depth to the certain range.
+        out = torch.clamp(out, min=self.decode_head.min_depth, max=self.decode_head.max_depth)
+        if rescale:
+            if size is None:
+                if img_metas is not None:
+                    size = img_metas[0]["ori_shape"][:2]
+                else:
+                    size = img.shape[2:]
+            out = resize(input=out, size=size, mode="bilinear", align_corners=self.align_corners)
+        return out
+
+    def _decode_head_forward_train(self, img, x, img_metas, depth_gt, **kwargs):
+        """Run forward function and calculate loss for decode head in
+        training."""
+        losses = dict()
+        loss_decode = self.decode_head.forward_train(img, x, img_metas, depth_gt, self.train_cfg, **kwargs)
+        losses.update(add_prefix(loss_decode, "decode"))
+        return losses
+
+    def _decode_head_forward_test(self, x, img_metas):
+        """Run forward function and calculate loss for decode head in
+        inference."""
+        depth_pred = self.decode_head.forward_test(x, img_metas, self.test_cfg)
+        return depth_pred
+
+    def forward_dummy(self, img):
+        """Dummy forward function."""
+        depth = self.encode_decode(img, None)
+
+        return depth
+
+    def forward_train(self, img, img_metas, depth_gt, **kwargs):
+        """Forward function for training.
+
+        Args:
+            img (Tensor): Input images.
+            img_metas (list[dict]): List of image info dict where each dict
+                has: 'img_shape', 'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `depth/datasets/pipelines/formatting.py:Collect`.
+            depth_gt (Tensor): Depth gt
+                used if the architecture supports depth estimation task.
+
+        Returns:
+            dict[str, Tensor]: a dictionary of loss components
+        """
+
+        x = self.extract_feat(img)
+
+        losses = dict()
+
+        # the last of x saves the info from neck
+        loss_decode = self._decode_head_forward_train(img, x, img_metas, depth_gt, **kwargs)
+
+        losses.update(loss_decode)
+
+        return losses
+
+    def whole_inference(self, img, img_meta, rescale, size=None):
+        """Inference with full image."""
+        depth_pred = self.encode_decode(img, img_meta, rescale, size=size)
+
+        return depth_pred
+
+    def slide_inference(self, img, img_meta, rescale):
+        """Inference by sliding-window with overlap.
+
+        If h_crop > h_img or w_crop > w_img, the small patch will be used to
+        decode without padding.
+        """
+
+        h_stride, w_stride = self.test_cfg.stride
+        h_crop, w_crop = self.test_cfg.crop_size
+        batch_size, _, h_img, w_img = img.size()
+        h_grids = max(h_img - h_crop + h_stride - 1, 0) // h_stride + 1
+        w_grids = max(w_img - w_crop + w_stride - 1, 0) // w_stride + 1
+        preds = img.new_zeros((batch_size, 1, h_img, w_img))
+        count_mat = img.new_zeros((batch_size, 1, h_img, w_img))
+        for h_idx in range(h_grids):
+            for w_idx in range(w_grids):
+                y1 = h_idx * h_stride
+                x1 = w_idx * w_stride
+                y2 = min(y1 + h_crop, h_img)
+                x2 = min(x1 + w_crop, w_img)
+                y1 = max(y2 - h_crop, 0)
+                x1 = max(x2 - w_crop, 0)
+                crop_img = img[:, :, y1:y2, x1:x2]
+                depth_pred = self.encode_decode(crop_img, img_meta, rescale)
+                preds += F.pad(depth_pred, (int(x1), int(preds.shape[3] - x2), int(y1), int(preds.shape[2] - y2)))
+
+                count_mat[:, :, y1:y2, x1:x2] += 1
+        assert (count_mat == 0).sum() == 0
+        if torch.onnx.is_in_onnx_export():
+            # cast count_mat to constant while exporting to ONNX
+            count_mat = torch.from_numpy(count_mat.cpu().detach().numpy()).to(device=img.device)
+        preds = preds / count_mat
+        return preds
+
+    def inference(self, img, img_meta, rescale, size=None):
+        """Inference with slide/whole style.
+
+        Args:
+            img (Tensor): The input image of shape (N, 3, H, W).
+            img_meta (dict): Image info dict where each dict has: 'img_shape',
+                'scale_factor', 'flip', and may also contain
+                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
+                For details on the values of these keys see
+                `depth/datasets/pipelines/formatting.py:Collect`.
+            rescale (bool): Whether rescale back to original shape.
+
+        Returns:
+            Tensor: The output depth map.
+        """
+
+        assert self.test_cfg.mode in ["slide", "whole"]
+        ori_shape = img_meta[0]["ori_shape"]
+        assert all(_["ori_shape"] == ori_shape for _ in img_meta)
+        if self.test_cfg.mode == "slide":
+            depth_pred = self.slide_inference(img, img_meta, rescale)
+        else:
+            depth_pred = self.whole_inference(img, img_meta, rescale, size=size)
+        output = depth_pred
+        flip = img_meta[0]["flip"]
+        if flip:
+            flip_direction = img_meta[0]["flip_direction"]
+            assert flip_direction in ["horizontal", "vertical"]
+            if flip_direction == "horizontal":
+                output = output.flip(dims=(3,))
+            elif flip_direction == "vertical":
+                output = output.flip(dims=(2,))
+
+        return output
+
+    def simple_test(self, img, img_meta, rescale=True):
+        """Simple test with single image."""
+        depth_pred = self.inference(img, img_meta, rescale)
+        if torch.onnx.is_in_onnx_export():
+            # our inference backend only support 4D output
+            depth_pred = depth_pred.unsqueeze(0)
+            return depth_pred
+        depth_pred = depth_pred.cpu().numpy()
+        # unravel batch dim
+        depth_pred = list(depth_pred)
+        return depth_pred
+
+    def aug_test(self, imgs, img_metas, rescale=True):
+        """Test with augmentations.
+
+        Only rescale=True is supported.
+        """
+        # aug_test rescale all imgs back to ori_shape for now
+        assert rescale
+        # to save memory, we get augmented depth logit inplace
+        depth_pred = self.inference(imgs[0], img_metas[0], rescale)
+        for i in range(1, len(imgs)):
+            cur_depth_pred = self.inference(imgs[i], img_metas[i], rescale, size=depth_pred.shape[-2:])
+            depth_pred += cur_depth_pred
+        depth_pred /= len(imgs)
+        depth_pred = depth_pred.cpu().numpy()
+        # unravel batch dim
+        depth_pred = list(depth_pred)
+        return depth_pred