Fix diff

src/transformers/models/dinov2/modeling_dinov2.py

@@ -292,37 +292,34 @@ 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(nn.Module):
     """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 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
+
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        return drop_path(hidden_states, self.drop_prob, self.training)
+        return self.drop_path(hidden_states, self.drop_prob, self.training)
 
     def extra_repr(self) -> str:
         return "p={}".format(self.drop_prob)

src/transformers/models/dinov2_with_registers/configuration_dinov2_with_registers.py

@@ -0,0 +1,172 @@
+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#               This file was automatically generated from <path_to_diff_file.py>.
+#         Do NOT edit this file manually as any edits will be overwritten by the generation of
+#         the file from the diff. If any change should be done, please apply the change to the
+#                                    diff.py file directly.
+#           🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2024 Meta Inc. 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.
+
+
+from transformers import PretrainedConfig
+
+from ...utils import (
+    logging,
+)
+from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices
+
+
+logger = logging.get_logger(__name__)
+
+
+class Dinov2WithRegistersConfig(BackboneConfigMixin, PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Dinov2WithRegistersModel`]. It is used to instantiate an
+    Dinov2WithRegisters model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Dinov2WithRegisters
+    [google/dinov2-with-registers-base-patch16-224](https://huggingface.co/google/dinov2-with-registers-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        mlp_ratio (`int`, *optional*, defaults to 4):
+            Ratio of the hidden size of the MLPs relative to the `hidden_size`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        layerscale_value (`float`, *optional*, defaults to 1.0):
+           Initial value to use for layer scale.
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            Stochastic depth rate per sample (when applied in the main path of residual layers).
+        use_swiglu_ffn (`bool`, *optional*, defaults to `False`):
+            Whether to use the SwiGLU feedforward neural network.
+        num_register_tokens (`int`, *optional*, defaults to 4):
+            Number of register tokens to use.
+        interpolate_antialias (`bool`, *optional*, defaults to `True`):
+            Whether to use antialiasing when interpolating the image patches.
+        interpolate_offset (`float`, *optional*, defaults to 0.0):
+            Offset to use when interpolating the image patches.
+        out_features (`List[str]`, *optional*):
+            If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc.
+            (depending on how many stages the model has). If unset and `out_indices` is set, will default to the
+            corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+        out_indices (`List[int]`, *optional*):
+            If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how
+            many stages the model has). If unset and `out_features` is set, will default to the corresponding stages.
+            If unset and `out_features` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+        apply_layernorm (`bool`, *optional*, defaults to `True`):
+            Whether to apply layer normalization to the feature maps in case the model is used as backbone.
+        reshape_hidden_states (`bool`, *optional*, defaults to `True`):
+            Whether to reshape the feature maps to 4D tensors of shape `(batch_size, hidden_size, height, width)` in
+            case the model is used as backbone. If `False`, the feature maps will be 3D tensors of shape `(batch_size,
+            seq_len, hidden_size)`.
+
+    Example:
+
+    ```python
+    >>> from transformers import Dinov2WithRegistersConfig, Dinov2WithRegistersModel
+
+    >>> # Initializing a Dinov2WithRegisters dinov2-with-registers-base-patch16-224 style configuration
+    >>> configuration = Dinov2WithRegistersConfig()
+
+    >>> # Initializing a model (with random weights) from the dinov2-with-registers-base-patch16-224 style configuration
+    >>> model = Dinov2WithRegistersModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "dinov2_with_registers"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        mlp_ratio=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-6,
+        image_size=224,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        layerscale_value=1.0,
+        drop_path_rate=0.0,
+        use_swiglu_ffn=False,
+        num_register_tokens=4,
+        interpolate_antialias=True,
+        interpolate_offset=0.0,
+        out_features=None,
+        out_indices=None,
+        apply_layernorm=True,
+        reshape_hidden_states=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.mlp_ratio = mlp_ratio
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.layerscale_value = layerscale_value
+        self.drop_path_rate = drop_path_rate
+        self.use_swiglu_ffn = use_swiglu_ffn
+        self.num_register_tokens = num_register_tokens
+        self.interpolate_antialias = interpolate_antialias
+        self.interpolate_offset = interpolate_offset
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, num_hidden_layers + 1)]
+        self._out_features, self._out_indices = get_aligned_output_features_output_indices(
+            out_features=out_features, out_indices=out_indices, stage_names=self.stage_names
+        )
+        self.apply_layernorm = apply_layernorm
+        self.reshape_hidden_states = reshape_hidden_states

src/transformers/models/dinov2_with_registers/diff_dinov2_with_registers.py

@@ -21,8 +21,15 @@
 from torch import nn
 
 from transformers import PretrainedConfig
-from transformers.models.dinov2.modeling_dinov2 import Dinov2ForImageClassification, Dinov2Model, Dinov2PatchEmbeddings
-
+from transformers.models.dinov2.modeling_dinov2 import (
+    Dinov2Backbone,
+    Dinov2Encoder,
+    Dinov2ForImageClassification,
+    Dinov2Model,
+    Dinov2PatchEmbeddings,
+)
+
+from ...modeling_outputs import BackboneOutput
 from ...utils import logging
 from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices
 
@@ -263,9 +270,107 @@ def forward(self, pixel_values: torch.Tensor, bool_masked_pos: Optional[torch.Te
         return embeddings
 
 
+class Dinov2WithRegistersEncoder(Dinov2Encoder):
+    pass
+
+
 class Dinov2WithRegistersModel(Dinov2Model):
     pass
 
 
 class Dinov2WithRegistersForImageClassification(Dinov2ForImageClassification):
     pass
+
+
+class Dinov2WithRegistersBackbone(Dinov2Backbone):
+    def __init__(self, config):
+        super().__init__(config)
+        super()._init_backbone(config)
+
+        self.num_register_tokens = config.num_register_tokens
+        self.num_features = [config.hidden_size for _ in range(config.num_hidden_layers + 1)]
+        self.embeddings = Dinov2WithRegistersEmbeddings(config)
+        self.encoder = Dinov2WithRegistersEncoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> Dinov2WithRegistersPatchEmbeddings:
+        return self.embeddings.patch_embeddings
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> BackboneOutput:
+        """
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, AutoBackbone
+        >>> import torch
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> processor = AutoImageProcessor.from_pretrained("facebook/dinov2_with_registers")
+        >>> model = AutoBackbone.from_pretrained(
+        ...     "facebook/dinov2_with_registers", out_features=["stage2", "stage5", "stage8", "stage11"]
+        ... )
+
+        >>> inputs = processor(image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> feature_maps = outputs.feature_maps
+        >>> list(feature_maps[-1].shape)
+        [1, 768, 16, 16]
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+
+        embedding_output = self.embeddings(pixel_values)
+
+        outputs = self.encoder(
+            embedding_output, output_hidden_states=True, output_attentions=output_attentions, return_dict=return_dict
+        )
+
+        hidden_states = outputs.hidden_states if return_dict else outputs[1]
+
+        feature_maps = ()
+        for stage, hidden_state in zip(self.stage_names, hidden_states):
+            if stage in self.out_features:
+                if self.config.apply_layernorm:
+                    hidden_state = self.layernorm(hidden_state)
+                if self.config.reshape_hidden_states:
+                    hidden_state = hidden_state[:, self.num_register_tokens + 1 :]
+                    # this was actually a bug in the original implementation that we copied here,
+                    # cause normally the order is height, width
+                    batch_size, _, height, width = pixel_values.shape
+                    patch_size = self.config.patch_size
+                    hidden_state = hidden_state.reshape(batch_size, height // patch_size, width // patch_size, -1)
+                    hidden_state = hidden_state.permute(0, 3, 1, 2).contiguous()
+                feature_maps += (hidden_state,)
+
+        if not return_dict:
+            if output_hidden_states:
+                output = (feature_maps,) + outputs[1:]
+            else:
+                output = (feature_maps,) + outputs[2:]
+            return output
+
+        return BackboneOutput(
+            feature_maps=feature_maps,
+            hidden_states=outputs.hidden_states if output_hidden_states else None,
+            attentions=outputs.attentions if output_attentions else None,
+        )