Add DETA (#20983)

* First draft

* Add initial draft of conversion script

* Convert all weights

* Fix config

* Add image processor

* Fix DetaImageProcessor

* Run make fix copies

* Remove timm dependency

* Fix dummy objects

* Improve loss function

* Remove conv_encoder attribute

* Update conversion scripts

* Improve postprocessing + docs

* Fix copied from statements

* Add tests

* Improve postprocessing

* Improve postprocessing

* Update READMEs

* More improvements

* Fix rebase

* Add is_torchvision_available

* Add torchvision dependency

* Fix typo and README

* Fix bug

* Add copied from

* Fix style

* Apply suggestions

* Fix thanks to @ydshieh

* Fix another dependency check

* Simplify image processor

* Add scipy

* Improve code

* Add threshold argument

* Fix bug

* Set default threshold

* Improve integration test

* Add another integration test

* Update setup.py

* Address review

* Improve deformable attention function

* Improve copied from

* Use relative imports

* Address review

* Replace assertions

* Address review

* Update dummies

* Remove dummies

* Address comments, update READMEs

* Remove custom kernel code

* Add image processor tests

* Add requires_backends

* Add minor comment

* Update scripts

* Update organization name

* Fix defaults, add doc tests

* Add id2label for object 365

* Fix tests

* Update task guide

src/transformers/models/deta/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace 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 typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {
+    "configuration_deta": ["DETA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetaConfig"],
+}
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_deta"] = ["DetaImageProcessor"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_deta"] = [
+        "DETA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DetaForObjectDetection",
+        "DetaModel",
+        "DetaPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_deta import DETA_PRETRAINED_CONFIG_ARCHIVE_MAP, DetaConfig
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_deta import DetaImageProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_deta import (
+            DETA_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DetaForObjectDetection,
+            DetaModel,
+            DetaPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/deta/configuration_deta.py

+# coding=utf-8
+# Copyright 2022 SenseTime 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.
+""" DETA model configuration"""
+
+import copy
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+DETA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "ut/deta": "https://huggingface.co/ut/deta/resolve/main/config.json",
+}
+
+
+class DetaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DetaModel`]. It is used to instantiate a DETA
+    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 DETA
+    [SenseTime/deformable-detr](https://huggingface.co/SenseTime/deformable-detr) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        backbone_config (`PretrainedConfig` or `dict`, *optional*, defaults to `ResNetConfig()`):
+            The configuration of the backbone model.
+        num_queries (`int`, *optional*, defaults to 900):
+            Number of object queries, i.e. detection slots. This is the maximal number of objects [`DetaModel`] can
+            detect in a single image. In case `two_stage` is set to `True`, we use `two_stage_num_proposals` instead.
+        d_model (`int`, *optional*, defaults to 256):
+            Dimension of the layers.
+        encoder_layers (`int`, *optional*, defaults to 6):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 6):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 2048):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        init_xavier_std (`float`, *optional*, defaults to 1):
+            The scaling factor used for the Xavier initialization gain in the HM Attention map module.
+        encoder_layerdrop: (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop: (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        auxiliary_loss (`bool`, *optional*, defaults to `False`):
+            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
+        position_embedding_type (`str`, *optional*, defaults to `"sine"`):
+            Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.
+        class_cost (`float`, *optional*, defaults to 1):
+            Relative weight of the classification error in the Hungarian matching cost.
+        bbox_cost (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.
+        giou_cost (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.
+        mask_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the Focal loss in the panoptic segmentation loss.
+        dice_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.
+        bbox_loss_coefficient (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 bounding box loss in the object detection loss.
+        giou_loss_coefficient (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss in the object detection loss.
+        eos_coefficient (`float`, *optional*, defaults to 0.1):
+            Relative classification weight of the 'no-object' class in the object detection loss.
+        num_feature_levels (`int`, *optional*, defaults to 5):
+            The number of input feature levels.
+        encoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the encoder.
+        decoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the decoder.
+        two_stage (`bool`, *optional*, defaults to `True`):
+            Whether to apply a two-stage deformable DETR, where the region proposals are also generated by a variant of
+            DETA, which are further fed into the decoder for iterative bounding box refinement.
+        two_stage_num_proposals (`int`, *optional*, defaults to 300):
+            The number of region proposals to be generated, in case `two_stage` is set to `True`.
+        with_box_refine (`bool`, *optional*, defaults to `True`):
+            Whether to apply iterative bounding box refinement, where each decoder layer refines the bounding boxes
+            based on the predictions from the previous layer.
+        focal_alpha (`float`, *optional*, defaults to 0.25):
+            Alpha parameter in the focal loss.
+
+    Examples:
+
+    ```python
+    >>> from transformers import DetaConfig, DetaModel
+
+    >>> # Initializing a DETA SenseTime/deformable-detr style configuration
+    >>> configuration = DetaConfig()
+
+    >>> # Initializing a model (with random weights) from the SenseTime/deformable-detr style configuration
+    >>> model = DetaModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "deta"
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(
+        self,
+        backbone_config=None,
+        num_queries=900,
+        max_position_embeddings=2048,
+        encoder_layers=6,
+        encoder_ffn_dim=2048,
+        encoder_attention_heads=8,
+        decoder_layers=6,
+        decoder_ffn_dim=1024,
+        decoder_attention_heads=8,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        is_encoder_decoder=True,
+        activation_function="relu",
+        d_model=256,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        init_xavier_std=1.0,
+        return_intermediate=True,
+        auxiliary_loss=False,
+        position_embedding_type="sine",
+        num_feature_levels=5,
+        encoder_n_points=4,
+        decoder_n_points=4,
+        two_stage=True,
+        two_stage_num_proposals=300,
+        with_box_refine=True,
+        assign_first_stage=True,
+        class_cost=1,
+        bbox_cost=5,
+        giou_cost=2,
+        mask_loss_coefficient=1,
+        dice_loss_coefficient=1,
+        bbox_loss_coefficient=5,
+        giou_loss_coefficient=2,
+        eos_coefficient=0.1,
+        focal_alpha=0.25,
+        **kwargs
+    ):
+        if backbone_config is None:
+            logger.info("`backbone_config` is `None`. Initializing the config with the default `ResNet` backbone.")
+            backbone_config = CONFIG_MAPPING["resnet"](out_features=["stage2", "stage3", "stage4"])
+        else:
+            if isinstance(backbone_config, dict):
+                backbone_model_type = backbone_config.pop("model_type")
+                config_class = CONFIG_MAPPING[backbone_model_type]
+                backbone_config = config_class.from_dict(backbone_config)
+
+        self.backbone_config = backbone_config
+        self.num_queries = num_queries
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.init_xavier_std = init_xavier_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.auxiliary_loss = auxiliary_loss
+        self.position_embedding_type = position_embedding_type
+        # deformable attributes
+        self.num_feature_levels = num_feature_levels
+        self.encoder_n_points = encoder_n_points
+        self.decoder_n_points = decoder_n_points
+        self.two_stage = two_stage
+        self.two_stage_num_proposals = two_stage_num_proposals
+        self.with_box_refine = with_box_refine
+        self.assign_first_stage = assign_first_stage
+        if two_stage is True and with_box_refine is False:
+            raise ValueError("If two_stage is True, with_box_refine must be True.")
+        # Hungarian matcher
+        self.class_cost = class_cost
+        self.bbox_cost = bbox_cost
+        self.giou_cost = giou_cost
+        # Loss coefficients
+        self.mask_loss_coefficient = mask_loss_coefficient
+        self.dice_loss_coefficient = dice_loss_coefficient
+        self.bbox_loss_coefficient = bbox_loss_coefficient
+        self.giou_loss_coefficient = giou_loss_coefficient
+        self.eos_coefficient = eos_coefficient
+        self.focal_alpha = focal_alpha
+        super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`]. Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["backbone_config"] = self.backbone_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/deta/convert_deta_resnet_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert DETA checkpoints from the original repository.
+
+URL: https://github.com/jozhang97/DETA/tree/master"""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+from huggingface_hub import cached_download, hf_hub_download, hf_hub_url
+from transformers import DetaConfig, DetaForObjectDetection, DetaImageProcessor
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_deta_config():
+    config = DetaConfig(
+        num_queries=900,
+        encoder_ffn_dim=2048,
+        decoder_ffn_dim=2048,
+        num_feature_levels=5,
+        assign_first_stage=True,
+        with_box_refine=True,
+        two_stage=True,
+    )
+
+    # set labels
+    config.num_labels = 91
+    repo_id = "huggingface/label-files"
+    filename = "coco-detection-id2label.json"
+    id2label = json.load(open(cached_download(hf_hub_url(repo_id, filename, repo_type="dataset")), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    return config
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config):
+    rename_keys = []
+
+    # stem
+    # fmt: off
+    rename_keys.append(("backbone.0.body.conv1.weight", "model.backbone.model.embedder.embedder.convolution.weight"))
+    rename_keys.append(("backbone.0.body.bn1.weight", "model.backbone.model.embedder.embedder.normalization.weight"))
+    rename_keys.append(("backbone.0.body.bn1.bias", "model.backbone.model.embedder.embedder.normalization.bias"))
+    rename_keys.append(("backbone.0.body.bn1.running_mean", "model.backbone.model.embedder.embedder.normalization.running_mean"))
+    rename_keys.append(("backbone.0.body.bn1.running_var", "model.backbone.model.embedder.embedder.normalization.running_var"))
+    # stages
+    for stage_idx in range(len(config.backbone_config.depths)):
+        for layer_idx in range(config.backbone_config.depths[stage_idx]):
+            # shortcut
+            if layer_idx == 0:
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.downsample.0.weight",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.shortcut.convolution.weight",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.downsample.1.weight",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.shortcut.normalization.weight",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.downsample.1.bias",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.shortcut.normalization.bias",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.downsample.1.running_mean",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.shortcut.normalization.running_mean",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.downsample.1.running_var",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.shortcut.normalization.running_var",
+                    )
+                )
+            # 3 convs
+            for i in range(3):
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.conv{i+1}.weight",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.layer.{i}.convolution.weight",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.bn{i+1}.weight",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.layer.{i}.normalization.weight",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.bn{i+1}.bias",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.layer.{i}.normalization.bias",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.bn{i+1}.running_mean",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.layer.{i}.normalization.running_mean",
+                    )
+                )
+                rename_keys.append(
+                    (
+                        f"backbone.0.body.layer{stage_idx + 1}.{layer_idx}.bn{i+1}.running_var",
+                        f"model.backbone.model.encoder.stages.{stage_idx}.layers.{layer_idx}.layer.{i}.normalization.running_var",
+                    )
+                )
+    # transformer encoder
+    for i in range(config.encoder_layers):
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.sampling_offsets.weight", f"model.encoder.layers.{i}.self_attn.sampling_offsets.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.sampling_offsets.bias", f"model.encoder.layers.{i}.self_attn.sampling_offsets.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.attention_weights.weight", f"model.encoder.layers.{i}.self_attn.attention_weights.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.attention_weights.bias", f"model.encoder.layers.{i}.self_attn.attention_weights.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.value_proj.weight", f"model.encoder.layers.{i}.self_attn.value_proj.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.value_proj.bias", f"model.encoder.layers.{i}.self_attn.value_proj.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.output_proj.weight", f"model.encoder.layers.{i}.self_attn.output_proj.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.output_proj.bias", f"model.encoder.layers.{i}.self_attn.output_proj.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm1.weight", f"model.encoder.layers.{i}.self_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm1.bias", f"model.encoder.layers.{i}.self_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear1.weight", f"model.encoder.layers.{i}.fc1.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear1.bias", f"model.encoder.layers.{i}.fc1.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear2.weight", f"model.encoder.layers.{i}.fc2.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear2.bias", f"model.encoder.layers.{i}.fc2.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm2.weight", f"model.encoder.layers.{i}.final_layer_norm.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm2.bias", f"model.encoder.layers.{i}.final_layer_norm.bias"))
+
+    # transformer decoder
+    for i in range(config.decoder_layers):
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.sampling_offsets.weight", f"model.decoder.layers.{i}.encoder_attn.sampling_offsets.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.sampling_offsets.bias", f"model.decoder.layers.{i}.encoder_attn.sampling_offsets.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.attention_weights.weight", f"model.decoder.layers.{i}.encoder_attn.attention_weights.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.attention_weights.bias", f"model.decoder.layers.{i}.encoder_attn.attention_weights.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.value_proj.weight", f"model.decoder.layers.{i}.encoder_attn.value_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.value_proj.bias", f"model.decoder.layers.{i}.encoder_attn.value_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.output_proj.weight", f"model.decoder.layers.{i}.encoder_attn.output_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.output_proj.bias", f"model.decoder.layers.{i}.encoder_attn.output_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm1.weight", f"model.decoder.layers.{i}.encoder_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm1.bias", f"model.decoder.layers.{i}.encoder_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.self_attn.out_proj.weight", f"model.decoder.layers.{i}.self_attn.out_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.self_attn.out_proj.bias", f"model.decoder.layers.{i}.self_attn.out_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm2.weight", f"model.decoder.layers.{i}.self_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm2.bias", f"model.decoder.layers.{i}.self_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear1.weight", f"model.decoder.layers.{i}.fc1.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear1.bias", f"model.decoder.layers.{i}.fc1.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear2.weight", f"model.decoder.layers.{i}.fc2.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear2.bias", f"model.decoder.layers.{i}.fc2.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm3.weight", f"model.decoder.layers.{i}.final_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm3.bias", f"model.decoder.layers.{i}.final_layer_norm.bias"))
+
+    # fmt: on
+
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+def read_in_decoder_q_k_v(state_dict, config):
+    # transformer decoder self-attention layers
+    hidden_size = config.d_model
+    for i in range(config.decoder_layers):
+        # read in weights + bias of input projection layer of self-attention
+        in_proj_weight = state_dict.pop(f"transformer.decoder.layers.{i}.self_attn.in_proj_weight")
+        in_proj_bias = state_dict.pop(f"transformer.decoder.layers.{i}.self_attn.in_proj_bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:hidden_size, :]
+        state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:hidden_size]
+        state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[
+            hidden_size : hidden_size * 2, :
+        ]
+        state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[hidden_size : hidden_size * 2]
+        state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-hidden_size:, :]
+        state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-hidden_size:]
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+
+    return im
+
+
+@torch.no_grad()
+def convert_deta_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub):
+    """
+    Copy/paste/tweak model's weights to our DETA structure.
+    """
+
+    # load config
+    config = get_deta_config()
+
+    # load original state dict
+    if model_name == "deta-resnet-50":
+        filename = "adet_checkpoint0011.pth"
+    elif model_name == "deta-resnet-50-24-epochs":
+        filename = "adet_2x_checkpoint0023.pth"
+    else:
+        raise ValueError(f"Model name {model_name} not supported")
+    checkpoint_path = hf_hub_download(repo_id="nielsr/deta-checkpoints", filename=filename)
+    state_dict = torch.load(checkpoint_path, map_location="cpu")["model"]
+
+    # rename keys
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+    read_in_decoder_q_k_v(state_dict, config)
+
+    # fix some prefixes
+    for key in state_dict.copy().keys():
+        if "transformer.decoder.class_embed" in key or "transformer.decoder.bbox_embed" in key:
+            val = state_dict.pop(key)
+            state_dict[key.replace("transformer.decoder", "model.decoder")] = val
+        if "input_proj" in key:
+            val = state_dict.pop(key)
+            state_dict["model." + key] = val
+        if "level_embed" in key or "pos_trans" in key or "pix_trans" in key or "enc_output" in key:
+            val = state_dict.pop(key)
+            state_dict[key.replace("transformer", "model")] = val
+
+    # finally, create HuggingFace model and load state dict
+    model = DetaForObjectDetection(config)
+    model.load_state_dict(state_dict)
+    model.eval()
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+
+    # load image processor
+    processor = DetaImageProcessor(format="coco_detection")
+
+    # verify our conversion on image
+    img = prepare_img()
+    encoding = processor(images=img, return_tensors="pt")
+    pixel_values = encoding["pixel_values"]
+    outputs = model(pixel_values.to(device))
+
+    # verify logits
+    if model_name == "deta-resnet-50":
+        expected_logits = torch.tensor(
+            [[-7.3978, -2.5406, -4.1668], [-8.2684, -3.9933, -3.8096], [-7.0515, -3.7973, -5.8516]]
+        )
+        expected_boxes = torch.tensor([[0.5043, 0.4973, 0.9998], [0.2542, 0.5489, 0.4748], [0.5490, 0.2765, 0.0570]])
+    elif model_name == "deta-resnet-50-24-epochs":
+        expected_logits = torch.tensor(
+            [[-7.1688, -2.4857, -4.8669], [-7.8630, -3.8154, -4.2674], [-7.2730, -4.1865, -5.5323]]
+        )
+        expected_boxes = torch.tensor([[0.5021, 0.4971, 0.9994], [0.2546, 0.5486, 0.4731], [0.1686, 0.1986, 0.2142]])
+
+    assert torch.allclose(outputs.logits[0, :3, :3], expected_logits.to(device), atol=1e-4)
+    assert torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes.to(device), atol=1e-4)
+    print("Everything ok!")
+
+    if pytorch_dump_folder_path:
+        # Save model and processor
+        logger.info(f"Saving PyTorch model and processor to {pytorch_dump_folder_path}...")
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    # Push to hub
+    if push_to_hub:
+        print("Pushing model and processor to hub...")
+        model.push_to_hub(f"jozhang97/{model_name}")
+        processor.push_to_hub(f"jozhang97/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        default="deta-resnet-50",
+        choices=["deta-resnet-50", "deta-resnet-50-24-epochs"],
+        help="Name of the model you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default=None,
+        type=str,
+        help="Path to the folder to output PyTorch model.",
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+    args = parser.parse_args()
+    convert_deta_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/deta/convert_deta_swin_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert DETA checkpoints from the original repository.
+
+URL: https://github.com/jozhang97/DETA/tree/master"""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+from huggingface_hub import cached_download, hf_hub_download, hf_hub_url
+from transformers import DetaConfig, DetaForObjectDetection, DetaImageProcessor, SwinConfig
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_deta_config(model_name):
+    backbone_config = SwinConfig(
+        embed_dim=192,
+        depths=(2, 2, 18, 2),
+        num_heads=(6, 12, 24, 48),
+        window_size=12,
+        out_features=["stage2", "stage3", "stage4"],
+    )
+
+    config = DetaConfig(
+        backbone_config=backbone_config,
+        num_queries=900,
+        encoder_ffn_dim=2048,
+        decoder_ffn_dim=2048,
+        num_feature_levels=5,
+        assign_first_stage=True,
+        with_box_refine=True,
+        two_stage=True,
+    )
+
+    # set labels
+    repo_id = "huggingface/label-files"
+    if "o365" in model_name:
+        num_labels = 366
+        filename = "object365-id2label.json"
+    else:
+        num_labels = 91
+        filename = "coco-detection-id2label.json"
+
+    config.num_labels = num_labels
+    id2label = json.load(open(cached_download(hf_hub_url(repo_id, filename, repo_type="dataset")), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    return config
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config):
+    rename_keys = []
+
+    # stem
+    # fmt: off
+    rename_keys.append(("backbone.0.body.patch_embed.proj.weight", "model.backbone.model.embeddings.patch_embeddings.projection.weight"))
+    rename_keys.append(("backbone.0.body.patch_embed.proj.bias", "model.backbone.model.embeddings.patch_embeddings.projection.bias"))
+    rename_keys.append(("backbone.0.body.patch_embed.norm.weight", "model.backbone.model.embeddings.norm.weight"))
+    rename_keys.append(("backbone.0.body.patch_embed.norm.bias", "model.backbone.model.embeddings.norm.bias"))
+    # stages
+    for i in range(len(config.backbone_config.depths)):
+        for j in range(config.backbone_config.depths[i]):
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.norm1.weight", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.layernorm_before.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.norm1.bias", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.layernorm_before.bias"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.attn.relative_position_bias_table", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.relative_position_bias_table"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.attn.relative_position_index", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.relative_position_index"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.attn.proj.weight", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.output.dense.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.attn.proj.bias", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.output.dense.bias"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.norm2.weight", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.layernorm_after.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.norm2.bias", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.layernorm_after.bias"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.mlp.fc1.weight", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.intermediate.dense.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.mlp.fc1.bias", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.intermediate.dense.bias"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.mlp.fc2.weight", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.output.dense.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.blocks.{j}.mlp.fc2.bias", f"model.backbone.model.encoder.layers.{i}.blocks.{j}.output.dense.bias"))
+
+        if i < 3:
+            rename_keys.append((f"backbone.0.body.layers.{i}.downsample.reduction.weight", f"model.backbone.model.encoder.layers.{i}.downsample.reduction.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.downsample.norm.weight", f"model.backbone.model.encoder.layers.{i}.downsample.norm.weight"))
+            rename_keys.append((f"backbone.0.body.layers.{i}.downsample.norm.bias", f"model.backbone.model.encoder.layers.{i}.downsample.norm.bias"))
+
+    rename_keys.append(("backbone.0.body.norm1.weight", "model.backbone.model.hidden_states_norms.stage2.weight"))
+    rename_keys.append(("backbone.0.body.norm1.bias", "model.backbone.model.hidden_states_norms.stage2.bias"))
+    rename_keys.append(("backbone.0.body.norm2.weight", "model.backbone.model.hidden_states_norms.stage3.weight"))
+    rename_keys.append(("backbone.0.body.norm2.bias", "model.backbone.model.hidden_states_norms.stage3.bias"))
+    rename_keys.append(("backbone.0.body.norm3.weight", "model.backbone.model.hidden_states_norms.stage4.weight"))
+    rename_keys.append(("backbone.0.body.norm3.bias", "model.backbone.model.hidden_states_norms.stage4.bias"))
+
+    # transformer encoder
+    for i in range(config.encoder_layers):
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.sampling_offsets.weight", f"model.encoder.layers.{i}.self_attn.sampling_offsets.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.sampling_offsets.bias", f"model.encoder.layers.{i}.self_attn.sampling_offsets.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.attention_weights.weight", f"model.encoder.layers.{i}.self_attn.attention_weights.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.attention_weights.bias", f"model.encoder.layers.{i}.self_attn.attention_weights.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.value_proj.weight", f"model.encoder.layers.{i}.self_attn.value_proj.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.value_proj.bias", f"model.encoder.layers.{i}.self_attn.value_proj.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.output_proj.weight", f"model.encoder.layers.{i}.self_attn.output_proj.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.self_attn.output_proj.bias", f"model.encoder.layers.{i}.self_attn.output_proj.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm1.weight", f"model.encoder.layers.{i}.self_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm1.bias", f"model.encoder.layers.{i}.self_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear1.weight", f"model.encoder.layers.{i}.fc1.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear1.bias", f"model.encoder.layers.{i}.fc1.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear2.weight", f"model.encoder.layers.{i}.fc2.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.linear2.bias", f"model.encoder.layers.{i}.fc2.bias"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm2.weight", f"model.encoder.layers.{i}.final_layer_norm.weight"))
+        rename_keys.append((f"transformer.encoder.layers.{i}.norm2.bias", f"model.encoder.layers.{i}.final_layer_norm.bias"))
+
+    # transformer decoder
+    for i in range(config.decoder_layers):
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.sampling_offsets.weight", f"model.decoder.layers.{i}.encoder_attn.sampling_offsets.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.sampling_offsets.bias", f"model.decoder.layers.{i}.encoder_attn.sampling_offsets.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.attention_weights.weight", f"model.decoder.layers.{i}.encoder_attn.attention_weights.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.attention_weights.bias", f"model.decoder.layers.{i}.encoder_attn.attention_weights.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.value_proj.weight", f"model.decoder.layers.{i}.encoder_attn.value_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.value_proj.bias", f"model.decoder.layers.{i}.encoder_attn.value_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.output_proj.weight", f"model.decoder.layers.{i}.encoder_attn.output_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.cross_attn.output_proj.bias", f"model.decoder.layers.{i}.encoder_attn.output_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm1.weight", f"model.decoder.layers.{i}.encoder_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm1.bias", f"model.decoder.layers.{i}.encoder_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.self_attn.out_proj.weight", f"model.decoder.layers.{i}.self_attn.out_proj.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.self_attn.out_proj.bias", f"model.decoder.layers.{i}.self_attn.out_proj.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm2.weight", f"model.decoder.layers.{i}.self_attn_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm2.bias", f"model.decoder.layers.{i}.self_attn_layer_norm.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear1.weight", f"model.decoder.layers.{i}.fc1.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear1.bias", f"model.decoder.layers.{i}.fc1.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear2.weight", f"model.decoder.layers.{i}.fc2.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.linear2.bias", f"model.decoder.layers.{i}.fc2.bias"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm3.weight", f"model.decoder.layers.{i}.final_layer_norm.weight"))
+        rename_keys.append((f"transformer.decoder.layers.{i}.norm3.bias", f"model.decoder.layers.{i}.final_layer_norm.bias"))
+
+    # fmt: on
+
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+# we split up the matrix of each encoder layer into queries, keys and values
+def read_in_swin_q_k_v(state_dict, backbone_config):
+    num_features = [int(backbone_config.embed_dim * 2**i) for i in range(len(backbone_config.depths))]
+    for i in range(len(backbone_config.depths)):
+        dim = num_features[i]
+        for j in range(backbone_config.depths[i]):
+            # fmt: off
+            # read in weights + bias of input projection layer (in original implementation, this is a single matrix + bias)
+            in_proj_weight = state_dict.pop(f"backbone.0.body.layers.{i}.blocks.{j}.attn.qkv.weight")
+            in_proj_bias = state_dict.pop(f"backbone.0.body.layers.{i}.blocks.{j}.attn.qkv.bias")
+            # next, add query, keys and values (in that order) to the state dict
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.query.weight"] = in_proj_weight[:dim, :]
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.query.bias"] = in_proj_bias[: dim]
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.key.weight"] = in_proj_weight[
+                dim : dim * 2, :
+            ]
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.key.bias"] = in_proj_bias[
+                dim : dim * 2
+            ]
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.value.weight"] = in_proj_weight[
+                -dim :, :
+            ]
+            state_dict[f"model.backbone.model.encoder.layers.{i}.blocks.{j}.attention.self.value.bias"] = in_proj_bias[-dim :]
+            # fmt: on
+
+
+def read_in_decoder_q_k_v(state_dict, config):
+    # transformer decoder self-attention layers
+    hidden_size = config.d_model
+    for i in range(config.decoder_layers):
+        # read in weights + bias of input projection layer of self-attention
+        in_proj_weight = state_dict.pop(f"transformer.decoder.layers.{i}.self_attn.in_proj_weight")
+        in_proj_bias = state_dict.pop(f"transformer.decoder.layers.{i}.self_attn.in_proj_bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:hidden_size, :]
+        state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:hidden_size]
+        state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[
+            hidden_size : hidden_size * 2, :
+        ]
+        state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[hidden_size : hidden_size * 2]
+        state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-hidden_size:, :]
+        state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-hidden_size:]
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+
+    return im
+
+
+@torch.no_grad()
+def convert_deta_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub):
+    """
+    Copy/paste/tweak model's weights to our DETA structure.
+    """
+
+    # load config
+    config = get_deta_config(model_name)
+
+    # load original state dict
+    if model_name == "deta-swin-large":
+        checkpoint_path = hf_hub_download(repo_id="nielsr/deta-checkpoints", filename="adet_swin_ft.pth")
+    elif model_name == "deta-swin-large-o365":
+        checkpoint_path = hf_hub_download(repo_id="jozhang97/deta-swin-l-o365", filename="deta_swin_pt_o365.pth")
+    else:
+        raise ValueError(f"Model name {model_name} not supported")
+
+    state_dict = torch.load(checkpoint_path, map_location="cpu")["model"]
+
+    # original state dict
+    for name, param in state_dict.items():
+        print(name, param.shape)
+
+    # rename keys
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+    read_in_swin_q_k_v(state_dict, config.backbone_config)
+    read_in_decoder_q_k_v(state_dict, config)
+
+    # fix some prefixes
+    for key in state_dict.copy().keys():
+        if "transformer.decoder.class_embed" in key or "transformer.decoder.bbox_embed" in key:
+            val = state_dict.pop(key)
+            state_dict[key.replace("transformer.decoder", "model.decoder")] = val
+        if "input_proj" in key:
+            val = state_dict.pop(key)
+            state_dict["model." + key] = val
+        if "level_embed" in key or "pos_trans" in key or "pix_trans" in key or "enc_output" in key:
+            val = state_dict.pop(key)
+            state_dict[key.replace("transformer", "model")] = val
+
+    # finally, create HuggingFace model and load state dict
+    model = DetaForObjectDetection(config)
+    model.load_state_dict(state_dict)
+    model.eval()
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+
+    # load image processor
+    processor = DetaImageProcessor(format="coco_detection")
+
+    # verify our conversion on image
+    img = prepare_img()
+    encoding = processor(images=img, return_tensors="pt")
+    pixel_values = encoding["pixel_values"]
+    outputs = model(pixel_values.to(device))
+
+    # verify logits
+    print("Logits:", outputs.logits[0, :3, :3])
+    print("Boxes:", outputs.pred_boxes[0, :3, :3])
+    if model_name == "deta-swin-large":
+        expected_logits = torch.tensor(
+            [[-7.6308, -2.8485, -5.3737], [-7.2037, -4.5505, -4.8027], [-7.2943, -4.2611, -4.6617]]
+        )
+        expected_boxes = torch.tensor([[0.4987, 0.4969, 0.9999], [0.2549, 0.5498, 0.4805], [0.5498, 0.2757, 0.0569]])
+    elif model_name == "deta-swin-large-o365":
+        expected_logits = torch.tensor(
+            [[-8.0122, -3.5720, -4.9717], [-8.1547, -3.6886, -4.6389], [-7.6610, -3.6194, -5.0134]]
+        )
+        expected_boxes = torch.tensor([[0.2523, 0.5549, 0.4881], [0.7715, 0.4149, 0.4601], [0.5503, 0.2753, 0.0575]])
+    assert torch.allclose(outputs.logits[0, :3, :3], expected_logits.to(device), atol=1e-4)
+    assert torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes.to(device), atol=1e-4)
+    print("Everything ok!")
+
+    if pytorch_dump_folder_path:
+        # Save model and processor
+        logger.info(f"Saving PyTorch model and processor to {pytorch_dump_folder_path}...")
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    # Push to hub
+    if push_to_hub:
+        print("Pushing model and processor to hub...")
+        model.push_to_hub(f"jozhang97/{model_name}")
+        processor.push_to_hub(f"jozhang97/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        default="deta-swin-large",
+        choices=["deta-swin-large", "deta-swin-large-o365"],
+        help="Name of the model you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default=None,
+        type=str,
+        help="Path to the folder to output PyTorch model.",
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+    args = parser.parse_args()
+    convert_deta_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/mask2former/modeling_mask2former.py

@@ -805,36 +805,42 @@ class Mask2FormerLoss(nn.Module):
         return num_masks_pt
 
 
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
 def multi_scale_deformable_attention(
     value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
-):
-    batch_size, _, num_attn_head, hidden_dim = value.shape
-    _, num_queries, num_attn_head, num_levels, num_points, _ = sampling_locations.shape
+) -> Tensor:
+    batch_size, _, num_heads, hidden_dim = value.shape
+    _, num_queries, num_heads, num_levels, num_points, _ = sampling_locations.shape
     value_list = value.split([height * width for height, width in value_spatial_shapes], dim=1)
     sampling_grids = 2 * sampling_locations - 1
     sampling_value_list = []
-
-    for idx, (height, width) in enumerate(value_spatial_shapes):
-        # batch_size, height*width, num_attn_head, hidden_dim -> batch_size*num_attn_head, hidden_dim, height, width
+    for level_id, (height, width) in enumerate(value_spatial_shapes):
+        # batch_size, height*width, num_heads, hidden_dim
+        # -> batch_size, height*width, num_heads*hidden_dim
+        # -> batch_size, num_heads*hidden_dim, height*width
+        # -> batch_size*num_heads, hidden_dim, height, width
         value_l_ = (
-            value_list[idx].flatten(2).transpose(1, 2).reshape(batch_size * num_attn_head, hidden_dim, height, width)
+            value_list[level_id].flatten(2).transpose(1, 2).reshape(batch_size * num_heads, hidden_dim, height, width)
         )
-        # (batch_size, num_queries, num_attn_head, num_points) -> (batch_size * num_attn_head, num_queries, num_points, 2)
-        sampling_grid_l_ = sampling_grids[:, :, :, idx].transpose(1, 2).flatten(0, 1)
-        # batch_size*num_attn_head, D_, num_queries, num_points
-        sampling_value_l_ = torch.nn.functional.grid_sample(
+        # batch_size, num_queries, num_heads, num_points, 2
+        # -> batch_size, num_heads, num_queries, num_points, 2
+        # -> batch_size*num_heads, num_queries, num_points, 2
+        sampling_grid_l_ = sampling_grids[:, :, :, level_id].transpose(1, 2).flatten(0, 1)
+        # batch_size*num_heads, hidden_dim, num_queries, num_points
+        sampling_value_l_ = nn.functional.grid_sample(
             value_l_, sampling_grid_l_, mode="bilinear", padding_mode="zeros", align_corners=False
         )
         sampling_value_list.append(sampling_value_l_)
-
-    # (batch_size, num_queries, num_attn_head, num_levels, num_points) -> (batch_size, num_attn_head, 1, num_queries, num_levels*num_points)
+    # (batch_size, num_queries, num_heads, num_levels, num_points)
+    # -> (batch_size, num_heads, num_queries, num_levels, num_points)
+    # -> (batch_size, num_heads, 1, num_queries, num_levels*num_points)
     attention_weights = attention_weights.transpose(1, 2).reshape(
-        batch_size * num_attn_head, 1, num_queries, num_levels * num_points
+        batch_size * num_heads, 1, num_queries, num_levels * num_points
     )
     output = (
         (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights)
         .sum(-1)
-        .view(batch_size, num_attn_head * hidden_dim, num_queries)
+        .view(batch_size, num_heads * hidden_dim, num_queries)
     )
     return output.transpose(1, 2).contiguous()
 

src/transformers/models/oneformer/modeling_oneformer.py

@@ -61,7 +61,8 @@ def _get_clones(module, N):
     return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
 
 
-def multiscale_deform_attn_core_pytorch(
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.multi_scale_deformable_attention
+def multi_scale_deformable_attention(
     value: Tensor, value_spatial_shapes: Tensor, sampling_locations: Tensor, attention_weights: Tensor
 ) -> Tensor:
     batch_size, _, num_heads, hidden_dim = value.shape
@@ -1044,8 +1045,8 @@ class OneFormerPixelDecoderEncoderMultiscaleDeformableAttention(nn.Module):
             )
         else:
             raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
-        # CPU
-        output = multiscale_deform_attn_core_pytorch(value, spatial_shapes, sampling_locations, attention_weights)
+        # PyTorch implementation
+        output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
         output = self.output_proj(output)
 
         return output, attention_weights

src/transformers/testing_utils.py

@@ -90,6 +90,7 @@ from .utils import (
     is_torch_tpu_available,
     is_torchaudio_available,
     is_torchdynamo_available,
+    is_torchvision_available,
     is_vision_available,
 )
 
@@ -307,6 +308,16 @@ def require_torch(test_case):
     return unittest.skipUnless(is_torch_available(), "test requires PyTorch")(test_case)
 
 
+def require_torchvision(test_case):
+    """
+    Decorator marking a test that requires Torchvision.
+
+    These tests are skipped when Torchvision isn't installed.
+
+    """
+    return unittest.skipUnless(is_torchvision_available(), "test requires Torchvision")(test_case)
+
+
 def require_torch_or_tf(test_case):
     """
     Decorator marking a test that requires PyTorch or TensorFlow.

src/transformers/utils/__init__.py

@@ -162,6 +162,7 @@ from .import_utils import (
     is_torchaudio_available,
     is_torchdistx_available,
     is_torchdynamo_available,
+    is_torchvision_available,
     is_training_run_on_sagemaker,
     is_vision_available,
     requires_backends,

src/transformers/utils/dummy_pt_objects.py

@@ -1991,6 +1991,30 @@ class DeiTPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+DETA_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class DetaForObjectDetection(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DetaModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class DetaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 DINAT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -136,6 +136,13 @@ class DeiTImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class DetaImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class DetrFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

src/transformers/utils/import_utils.py

@@ -288,6 +288,10 @@ def is_torch_available():
     return _torch_available
 
 
+def is_torchvision_available():
+    return importlib.util.find_spec("torchvision") is not None
+
+
 def is_pyctcdecode_available():
     return _pyctcdecode_available
 
@@ -803,6 +807,14 @@ installation page: https://pytorch.org/get-started/locally/ and follow the ones
 Please note that you may need to restart your runtime after installation.
 """
 
+
+# docstyle-ignore
+TORCHVISION_IMPORT_ERROR = """
+{0} requires the Torchvision library but it was not found in your environment. Checkout the instructions on the
+installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment.
+Please note that you may need to restart your runtime after installation.
+"""
+
 # docstyle-ignore
 PYTORCH_IMPORT_ERROR_WITH_TF = """
 {0} requires the PyTorch library but it was not found in your environment.
@@ -1009,6 +1021,7 @@ BACKENDS_MAPPING = OrderedDict(
         ("natten", (is_natten_available, NATTEN_IMPORT_ERROR)),
         ("tokenizers", (is_tokenizers_available, TOKENIZERS_IMPORT_ERROR)),
         ("torch", (is_torch_available, PYTORCH_IMPORT_ERROR)),
+        ("torchvision", (is_torchvision_available, TORCHVISION_IMPORT_ERROR)),
         ("vision", (is_vision_available, VISION_IMPORT_ERROR)),
         ("scipy", (is_scipy_available, SCIPY_IMPORT_ERROR)),
         ("accelerate", (is_accelerate_available, ACCELERATE_IMPORT_ERROR)),

tests/models/deta/test_image_processing_deta.py

+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# 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.
+
+
+import json
+import pathlib
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision, slow
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingSavingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import DetaImageProcessor
+
+
+class DetaImageProcessingTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_pad=True,
+    ):
+        # by setting size["longest_edge"] > max_resolution we're effectively not testing this :p
+        size = size if size is not None else {"shortest_edge": 18, "longest_edge": 1333}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_pad = do_pad
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_rescale": self.do_rescale,
+            "rescale_factor": self.rescale_factor,
+            "do_pad": self.do_pad,
+        }
+
+    def get_expected_values(self, image_inputs, batched=False):
+        """
+        This function computes the expected height and width when providing images to DetaImageProcessor,
+        assuming do_resize is set to True with a scalar size.
+        """
+        if not batched:
+            image = image_inputs[0]
+            if isinstance(image, Image.Image):
+                w, h = image.size
+            else:
+                h, w = image.shape[1], image.shape[2]
+            if w < h:
+                expected_height = int(self.size["shortest_edge"] * h / w)
+                expected_width = self.size["shortest_edge"]
+            elif w > h:
+                expected_height = self.size["shortest_edge"]
+                expected_width = int(self.size["shortest_edge"] * w / h)
+            else:
+                expected_height = self.size["shortest_edge"]
+                expected_width = self.size["shortest_edge"]
+
+        else:
+            expected_values = []
+            for image in image_inputs:
+                expected_height, expected_width = self.get_expected_values([image])
+                expected_values.append((expected_height, expected_width))
+            expected_height = max(expected_values, key=lambda item: item[0])[0]
+            expected_width = max(expected_values, key=lambda item: item[1])[1]
+
+        return expected_height, expected_width
+
+
+@require_torch
+@require_vision
+class DetaImageProcessingTest(ImageProcessingSavingTestMixin, unittest.TestCase):
+
+    image_processing_class = DetaImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.image_processor_tester = DetaImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        self.assertTrue(hasattr(image_processing, "image_mean"))
+        self.assertTrue(hasattr(image_processing, "image_std"))
+        self.assertTrue(hasattr(image_processing, "do_normalize"))
+        self.assertTrue(hasattr(image_processing, "do_resize"))
+        self.assertTrue(hasattr(image_processing, "do_rescale"))
+        self.assertTrue(hasattr(image_processing, "do_pad"))
+        self.assertTrue(hasattr(image_processing, "size"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        image_processor = self.image_processing_class.from_dict(self.image_processor_dict)
+        self.assertEqual(image_processor.size, {"shortest_edge": 18, "longest_edge": 1333})
+        self.assertEqual(image_processor.do_pad, True)
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PIL images
+        image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
+
+        self.assertEqual(
+            encoded_images.shape,
+            (1, self.image_processor_tester.num_channels, expected_height, expected_width),
+        )
+
+        # Test batched
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
+
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random numpy tensors
+        image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
+
+        self.assertEqual(
+            encoded_images.shape,
+            (1, self.image_processor_tester.num_channels, expected_height, expected_width),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
+
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_class(**self.image_processor_dict)
+        # create random PyTorch tensors
+        image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
+
+        self.assertEqual(
+            encoded_images.shape,
+            (1, self.image_processor_tester.num_channels, expected_height, expected_width),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+
+        expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
+
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.image_processor_tester.batch_size,
+                self.image_processor_tester.num_channels,
+                expected_height,
+                expected_width,
+            ),
+        )
+
+    def test_equivalence_pad_and_create_pixel_mask(self):
+        # Initialize image_processings
+        image_processing_1 = self.image_processing_class(**self.image_processor_dict)
+        image_processing_2 = self.image_processing_class(do_resize=False, do_normalize=False, do_rescale=False)
+
+        # create random PyTorch tensors
+        image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test whether the method "pad_and_return_pixel_mask" and calling the image processor return the same tensors
+        encoded_images_with_method = image_processing_1.pad_and_create_pixel_mask(image_inputs, return_tensors="pt")
+        encoded_images = image_processing_2(image_inputs, return_tensors="pt")
+
+        self.assertTrue(
+            torch.allclose(encoded_images_with_method["pixel_values"], encoded_images["pixel_values"], atol=1e-4)
+        )
+        self.assertTrue(
+            torch.allclose(encoded_images_with_method["pixel_mask"], encoded_images["pixel_mask"], atol=1e-4)
+        )
+
+    @slow
+    def test_call_pytorch_with_coco_detection_annotations(self):
+        # prepare image and target
+        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+        with open("./tests/fixtures/tests_samples/COCO/coco_annotations.txt", "r") as f:
+            target = json.loads(f.read())
+
+        target = {"image_id": 39769, "annotations": target}
+
+        # encode them
+        image_processing = DetaImageProcessor()
+        encoding = image_processing(images=image, annotations=target, return_tensors="pt")
+
+        # verify pixel values
+        expected_shape = torch.Size([1, 3, 800, 1066])
+        self.assertEqual(encoding["pixel_values"].shape, expected_shape)
+
+        expected_slice = torch.tensor([0.2796, 0.3138, 0.3481])
+        self.assertTrue(torch.allclose(encoding["pixel_values"][0, 0, 0, :3], expected_slice, atol=1e-4))
+
+        # verify area
+        expected_area = torch.tensor([5887.9600, 11250.2061, 489353.8438, 837122.7500, 147967.5156, 165732.3438])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["area"], expected_area))
+        # verify boxes
+        expected_boxes_shape = torch.Size([6, 4])
+        self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape)
+        expected_boxes_slice = torch.tensor([0.5503, 0.2765, 0.0604, 0.2215])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3))
+        # verify image_id
+        expected_image_id = torch.tensor([39769])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["image_id"], expected_image_id))
+        # verify is_crowd
+        expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd))
+        # verify class_labels
+        expected_class_labels = torch.tensor([75, 75, 63, 65, 17, 17])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels))
+        # verify orig_size
+        expected_orig_size = torch.tensor([480, 640])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size))
+        # verify size
+        expected_size = torch.tensor([800, 1066])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["size"], expected_size))
+
+    @slow
+    def test_call_pytorch_with_coco_panoptic_annotations(self):
+        # prepare image, target and masks_path
+        image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+        with open("./tests/fixtures/tests_samples/COCO/coco_panoptic_annotations.txt", "r") as f:
+            target = json.loads(f.read())
+
+        target = {"file_name": "000000039769.png", "image_id": 39769, "segments_info": target}
+
+        masks_path = pathlib.Path("./tests/fixtures/tests_samples/COCO/coco_panoptic")
+
+        # encode them
+        image_processing = DetaImageProcessor(format="coco_panoptic")
+        encoding = image_processing(images=image, annotations=target, masks_path=masks_path, return_tensors="pt")
+
+        # verify pixel values
+        expected_shape = torch.Size([1, 3, 800, 1066])
+        self.assertEqual(encoding["pixel_values"].shape, expected_shape)
+
+        expected_slice = torch.tensor([0.2796, 0.3138, 0.3481])
+        self.assertTrue(torch.allclose(encoding["pixel_values"][0, 0, 0, :3], expected_slice, atol=1e-4))
+
+        # verify area
+        expected_area = torch.tensor([147979.6875, 165527.0469, 484638.5938, 11292.9375, 5879.6562, 7634.1147])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["area"], expected_area))
+        # verify boxes
+        expected_boxes_shape = torch.Size([6, 4])
+        self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape)
+        expected_boxes_slice = torch.tensor([0.2625, 0.5437, 0.4688, 0.8625])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3))
+        # verify image_id
+        expected_image_id = torch.tensor([39769])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["image_id"], expected_image_id))
+        # verify is_crowd
+        expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd))
+        # verify class_labels
+        expected_class_labels = torch.tensor([17, 17, 63, 75, 75, 93])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels))
+        # verify masks
+        expected_masks_sum = 822873
+        self.assertEqual(encoding["labels"][0]["masks"].sum().item(), expected_masks_sum)
+        # verify orig_size
+        expected_orig_size = torch.tensor([480, 640])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size))
+        # verify size
+        expected_size = torch.tensor([800, 1066])
+        self.assertTrue(torch.allclose(encoding["labels"][0]["size"], expected_size))

utils/check_repo.py

@@ -53,6 +53,8 @@ PRIVATE_MODELS = [
 # Being in this list is an exception and should **not** be the rule.
 IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     # models to ignore for not tested
+    "DetaEncoder",  # Building part of bigger (tested) model.
+    "DetaDecoder",  # Building part of bigger (tested) model.
     "GraphormerEncoder",  # Building part of bigger (tested) model.
     "GraphormerDecoderHead",  # Building part of bigger (tested) model.
     "CLIPSegDecoder",  # Building part of bigger (tested) model.

utils/documentation_tests.txt

@@ -67,6 +67,8 @@ src/transformers/models/deformable_detr/modeling_deformable_detr.py
 src/transformers/models/deit/configuration_deit.py
 src/transformers/models/deit/modeling_deit.py
 src/transformers/models/deit/modeling_tf_deit.py
+src/transformers/models/deta/configuration_deta.py
+src/transformers/models/deta/modeling_deta.py
 src/transformers/models/detr/configuration_detr.py
 src/transformers/models/detr/modeling_detr.py
 src/transformers/models/dinat/configuration_dinat.py