Add DINOv2 (#24016)

* First draft

* More improvements

* Convert patch embedding layer

* Convert all weights

* Make conversion work

* Improve conversion script

* Fix style

* Make all tests pass

* Add image processor to auto mapping

* Add swiglu ffn

* Add image processor to conversion script

* Fix conversion of giant model

* Fix documentation

* Fix style

* Fix tests

* Address comments

* Address more comments

* Remove unused arguments

* Remove more arguments

* Rename parameters

* Include mask token

* Address comments

* Add docstring

* Transfer checkpoints

* Empty commit

src/transformers/models/dinov2/modeling_dinov2.py

+# coding=utf-8
+# Copyright 2023 Meta AI and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch DINOv2 model."""
+
+
+import collections.abc
+import math
+from typing import Dict, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPooling,
+    ImageClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+)
+from .configuration_dinov2 import Dinov2Config
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "Dinov2Config"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "facebook/dinov2-base"
+_EXPECTED_OUTPUT_SHAPE = [1, 197, 768]
+
+# Image classification docstring
+_IMAGE_CLASS_CHECKPOINT = "facebook/dinov2-base-patch16-224"
+_IMAGE_CLASS_EXPECTED_OUTPUT = "Egyptian cat"
+
+
+DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/dinov2-base",
+    # See all DINOv2 models at https://huggingface.co/models?filter=dinov2
+]
+
+
+class Dinov2Embeddings(nn.Module):
+    """
+    Construct the CLS token, mask token, position and patch embeddings.
+    """
+
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+
+        self.cls_token = nn.Parameter(torch.randn(1, 1, config.hidden_size))
+        self.mask_token = nn.Parameter(torch.zeros(1, config.hidden_size))
+        self.patch_embeddings = Dinov2PatchEmbeddings(config)
+        num_patches = self.patch_embeddings.num_patches
+        self.position_embeddings = nn.Parameter(torch.randn(1, num_patches + 1, config.hidden_size))
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
+        resolution images.
+
+        Source:
+        https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
+        """
+
+        num_patches = embeddings.shape[1] - 1
+        num_positions = self.position_embeddings.shape[1] - 1
+        if num_patches == num_positions and height == width:
+            return self.position_embeddings
+        class_pos_embed = self.position_embeddings[:, 0]
+        patch_pos_embed = self.position_embeddings[:, 1:]
+        dim = embeddings.shape[-1]
+        height = height // self.config.patch_size
+        width = width // self.config.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        height, width = height + 0.1, width + 0.1
+        patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            scale_factor=(height / math.sqrt(num_positions), width / math.sqrt(num_positions)),
+            mode="bicubic",
+            align_corners=False,
+        )
+        if int(height) != patch_pos_embed.shape[-2] or int(width) != patch_pos_embed.shape[-1]:
+            raise ValueError("Width or height does not match with the interpolated position embeddings")
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def forward(self, pixel_values: torch.Tensor, bool_masked_pos: torch.Tensor) -> torch.Tensor:
+        batch_size, _, height, width = pixel_values.shape
+        embeddings = self.patch_embeddings(pixel_values)
+
+        if bool_masked_pos is not None:
+            embeddings = torch.where(
+                bool_masked_pos.unsqueeze(-1), self.mask_token.to(embeddings.dtype).unsqueeze(0), embeddings
+            )
+
+        # add the [CLS] token to the embedded patch tokens
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
+        embeddings = torch.cat((cls_tokens, embeddings), dim=1)
+
+        # add positional encoding to each token
+        embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+
+        embeddings = self.dropout(embeddings)
+
+        return embeddings
+
+
+class Dinov2PatchEmbeddings(nn.Module):
+    """
+    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
+    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
+    Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        image_size, patch_size = config.image_size, config.patch_size
+        num_channels, hidden_size = config.num_channels, config.hidden_size
+
+        image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
+        patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.num_patches = num_patches
+
+        self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        num_channels = pixel_values.shape[1]
+        if num_channels != self.num_channels:
+            raise ValueError(
+                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
+                f" Expected {self.num_channels} but got {num_channels}."
+            )
+        embeddings = self.projection(pixel_values).flatten(2).transpose(1, 2)
+        return embeddings
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfAttention with ViT->Dinov2
+class Dinov2SelfAttention(nn.Module):
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
+                f"heads {config.num_attention_heads}."
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self, hidden_states, head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->Dinov2
+class Dinov2SelfOutput(nn.Module):
+    """
+    The residual connection is defined in Dinov2Layer instead of here (as is the case with other models), due to the
+    layernorm applied before each block.
+    """
+
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->Dinov2
+class Dinov2Attention(nn.Module):
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+        self.attention = Dinov2SelfAttention(config)
+        self.output = Dinov2SelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads: Set[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_outputs = self.attention(hidden_states, head_mask, output_attentions)
+
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class Dinov2LayerScale(nn.Module):
+    def __init__(self, config) -> None:
+        super().__init__()
+        self.lambda1 = nn.Parameter(config.layerscale_value * torch.ones(config.hidden_size))
+
+    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        return hidden_state * self.lambda1
+
+
+# Copied from transformers.models.beit.modeling_beit.drop_path
+def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:
+    """
+    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
+    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
+    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
+    argument.
+    """
+    if drop_prob == 0.0 or not training:
+        return input
+    keep_prob = 1 - drop_prob
+    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)
+    random_tensor.floor_()  # binarize
+    output = input.div(keep_prob) * random_tensor
+    return output
+
+
+# Copied from transformers.models.beit.modeling_beit.BeitDropPath
+class Dinov2DropPath:
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob: Optional[float] = None) -> None:
+        super().__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return drop_path(hidden_states, self.drop_prob, self.training)
+
+    def extra_repr(self) -> str:
+        return "p={}".format(self.drop_prob)
+
+
+class Dinov2MLP(nn.Module):
+    def __init__(self, config) -> None:
+        super().__init__()
+        in_features = out_features = config.hidden_size
+        hidden_features = int(config.hidden_size * config.mlp_ratio)
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=True)
+        if isinstance(config.hidden_act, str):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=True)
+
+    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        hidden_state = self.fc1(hidden_state)
+        hidden_state = self.activation(hidden_state)
+        hidden_state = self.fc2(hidden_state)
+        return hidden_state
+
+
+class Dinov2SwiGLUFFN(nn.Module):
+    def __init__(self, config) -> None:
+        super().__init__()
+        in_features = out_features = config.hidden_size
+        hidden_features = int(config.hidden_size * config.mlp_ratio)
+        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
+
+        self.weights_in = nn.Linear(in_features, 2 * hidden_features, bias=True)
+        self.weights_out = nn.Linear(hidden_features, out_features, bias=True)
+
+    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
+        hidden_state = self.weights_in(hidden_state)
+        x1, x2 = hidden_state.chunk(2, dim=-1)
+        hidden = nn.functional.silu(x1) * x2
+        return self.weights_out(hidden)
+
+
+class Dinov2Layer(nn.Module):
+    """This corresponds to the Block class in the original implementation."""
+
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+
+        self.norm1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = Dinov2Attention(config)
+        self.layer_scale1 = Dinov2LayerScale(config)
+        self.drop_path1 = Dinov2DropPath(config.drop_path_rate) if config.drop_path_rate > 0.0 else nn.Identity()
+
+        self.norm2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        if config.use_swiglu_ffn:
+            self.mlp = Dinov2SwiGLUFFN(config)
+        else:
+            self.mlp = Dinov2MLP(config)
+        self.layer_scale2 = Dinov2LayerScale(config)
+        self.drop_path2 = Dinov2DropPath(config.drop_path_rate) if config.drop_path_rate > 0.0 else nn.Identity()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_attention_outputs = self.attention(
+            self.norm1(hidden_states),  # in Dinov2, layernorm is applied before self-attention
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+
+        attention_output = self.layer_scale1(attention_output)
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # first residual connection
+        hidden_states = attention_output + hidden_states
+
+        # in Dinov2, layernorm is also applied after self-attention
+        layer_output = self.norm2(hidden_states)
+        layer_output = self.mlp(layer_output)
+        layer_output = self.layer_scale2(layer_output)
+
+        # second residual connection
+        layer_output = layer_output + hidden_states
+
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTEncoder with ViT->Dinov2
+class Dinov2Encoder(nn.Module):
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([Dinov2Layer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class Dinov2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = Dinov2Config
+    base_model_prefix = "dinov2"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
+            # `trunc_normal_cpu` not implemented in `half` issues
+            module.weight.data = nn.init.trunc_normal_(
+                module.weight.data.to(torch.float32), mean=0.0, std=self.config.initializer_range
+            ).to(module.weight.dtype)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, Dinov2Embeddings):
+            module.position_embeddings.data = nn.init.trunc_normal_(
+                module.position_embeddings.data.to(torch.float32),
+                mean=0.0,
+                std=self.config.initializer_range,
+            ).to(module.position_embeddings.dtype)
+
+            module.cls_token.data = nn.init.trunc_normal_(
+                module.cls_token.data.to(torch.float32),
+                mean=0.0,
+                std=self.config.initializer_range,
+            ).to(module.cls_token.dtype)
+
+    def _set_gradient_checkpointing(self, module: Dinov2Encoder, value: bool = False) -> None:
+        if isinstance(module, Dinov2Encoder):
+            module.gradient_checkpointing = value
+
+
+DINOV2_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`Dinov2Config`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+DINOV2_BASE_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
+            [`BitImageProcessor.preprocess`] for details.
+
+        bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, sequence_length)`):
+            Boolean masked positions. Indicates which patches are masked (1) and which aren't (0). Only relevant for
+            pre-training.
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+DINOV2_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
+            [`BitImageProcessor.preprocess`] for details.
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare DINOv2 Model transformer outputting raw hidden-states without any specific head on top.",
+    DINOV2_START_DOCSTRING,
+)
+class Dinov2Model(Dinov2PreTrainedModel):
+    def __init__(self, config: Dinov2Config, add_pooling_layer: bool = True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = Dinov2Embeddings(config)
+        self.encoder = Dinov2Encoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pooler = Dinov2Pooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> Dinov2PatchEmbeddings:
+        return self.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(DINOV2_BASE_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+        modality="vision",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        bool_masked_pos: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(pixel_values, bool_masked_pos=bool_masked_pos)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        sequence_output = self.layernorm(sequence_output)
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            head_outputs = (sequence_output, pooled_output) if pooled_output is not None else (sequence_output,)
+            return head_outputs + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTPooler with ViT->Dinov2
+class Dinov2Pooler(nn.Module):
+    def __init__(self, config: Dinov2Config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+@add_start_docstrings(
+    """
+    Dinov2 Model transformer with an image classification head on top (a linear layer on top of the final hidden state
+    of the [CLS] token) e.g. for ImageNet.
+    """,
+    DINOV2_START_DOCSTRING,
+)
+class Dinov2ForImageClassification(Dinov2PreTrainedModel):
+    def __init__(self, config: Dinov2Config) -> None:
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.dinov2 = Dinov2Model(config, add_pooling_layer=False)
+
+        # Classifier head
+        self.classifier = (
+            nn.Linear(config.hidden_size * 2, config.num_labels) if config.num_labels > 0 else nn.Identity()
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(DINOV2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_IMAGE_CLASS_CHECKPOINT,
+        output_type=ImageClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[tuple, ImageClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.dinov2(
+            pixel_values,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]  # batch_size, sequence_length, hidden_size
+
+        cls_token = sequence_output[:, 0]
+        patch_tokens = sequence_output[:, 1:]
+
+        linear_input = torch.cat([cls_token, patch_tokens.mean(dim=1)], dim=1)
+
+        logits = self.classifier(linear_input)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return ImageClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -2540,6 +2540,30 @@ class DinatPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class Dinov2ForImageClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Dinov2Model(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Dinov2PreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/dinov2/test_modeling_dinov2.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch Dinov2 model. """
+
+
+import inspect
+import unittest
+
+from transformers import Dinov2Config
+from transformers.testing_utils import (
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import cached_property, is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import Dinov2ForImageClassification, Dinov2Model
+    from transformers.models.dinov2.modeling_dinov2 import DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoImageProcessor
+
+
+class Dinov2ModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        type_sequence_label_size=10,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        # in Dinov2, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return Dinov2Config(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = Dinov2Model(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.type_sequence_label_size
+        model = Dinov2ForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+        # test greyscale images
+        config.num_channels = 1
+        model = Dinov2ForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+
+        pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
+        result = model(pixel_values)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            pixel_values,
+            labels,
+        ) = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class Dinov2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as Dinov2 does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (Dinov2Model, Dinov2ForImageClassification) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {"feature-extraction": Dinov2Model, "image-classification": Dinov2ForImageClassification}
+        if is_torch_available()
+        else {}
+    )
+    fx_compatible = False
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = Dinov2ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Dinov2Config, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="Dinov2 does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_image_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = Dinov2Model.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@require_torch
+@require_vision
+class Dinov2ModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return AutoImageProcessor.from_pretrained("facebook/dinov2-base") if is_vision_available() else None
+
+    @slow
+    def test_inference_no_head(self):
+        model = Dinov2Model.from_pretrained("facebook/dinov2-base").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the last hidden states
+        expected_shape = torch.Size((1, 257, 768))
+        self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[-2.1747, -0.4729, 1.0936], [-3.2780, -0.8269, -0.9210], [-2.9129, 1.1284, -0.7306]],
+            device=torch_device,
+        )
+        self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))