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/deta/image_processing_deta.py

+# coding=utf-8
+# Copyright 2022 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.
+"""Image processor class for Deformable DETR."""
+
+import pathlib
+import warnings
+from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple, Union
+
+import numpy as np
+
+from ...feature_extraction_utils import BatchFeature
+from ...image_processing_utils import BaseImageProcessor, get_size_dict
+from ...image_transforms import (
+    PaddingMode,
+    center_to_corners_format,
+    corners_to_center_format,
+    normalize,
+    pad,
+    rescale,
+    resize,
+    rgb_to_id,
+    to_channel_dimension_format,
+)
+from ...image_utils import (
+    IMAGENET_DEFAULT_MEAN,
+    IMAGENET_DEFAULT_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_batched,
+    to_numpy_array,
+    valid_coco_detection_annotations,
+    valid_coco_panoptic_annotations,
+    valid_images,
+)
+from ...utils import (
+    is_flax_available,
+    is_jax_tensor,
+    is_tf_available,
+    is_tf_tensor,
+    is_torch_available,
+    is_torch_tensor,
+    is_torchvision_available,
+    is_vision_available,
+)
+from ...utils.generic import ExplicitEnum, TensorType
+
+
+if is_torch_available():
+    import torch
+
+if is_torchvision_available():
+    from torchvision.ops.boxes import batched_nms
+
+if is_vision_available():
+    import PIL
+
+
+class AnnotionFormat(ExplicitEnum):
+    COCO_DETECTION = "coco_detection"
+    COCO_PANOPTIC = "coco_panoptic"
+
+
+SUPPORTED_ANNOTATION_FORMATS = (AnnotionFormat.COCO_DETECTION, AnnotionFormat.COCO_PANOPTIC)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_size_with_aspect_ratio
+def get_size_with_aspect_ratio(image_size, size, max_size=None) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image size and the desired output size.
+
+    Args:
+        image_size (`Tuple[int, int]`):
+            The input image size.
+        size (`int`):
+            The desired output size.
+        max_size (`int`, *optional*):
+            The maximum allowed output size.
+    """
+    height, width = image_size
+    if max_size is not None:
+        min_original_size = float(min((height, width)))
+        max_original_size = float(max((height, width)))
+        if max_original_size / min_original_size * size > max_size:
+            size = int(round(max_size * min_original_size / max_original_size))
+
+    if (height <= width and height == size) or (width <= height and width == size):
+        return height, width
+
+    if width < height:
+        ow = size
+        oh = int(size * height / width)
+    else:
+        oh = size
+        ow = int(size * width / height)
+    return (oh, ow)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_resize_output_image_size
+def get_resize_output_image_size(
+    input_image: np.ndarray, size: Union[int, Tuple[int, int], List[int]], max_size: Optional[int] = None
+) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image size and the desired output size. If the desired output size
+    is a tuple or list, the output image size is returned as is. If the desired output size is an integer, the output
+    image size is computed by keeping the aspect ratio of the input image size.
+
+    Args:
+        image_size (`Tuple[int, int]`):
+            The input image size.
+        size (`int`):
+            The desired output size.
+        max_size (`int`, *optional*):
+            The maximum allowed output size.
+    """
+    image_size = get_image_size(input_image)
+    if isinstance(size, (list, tuple)):
+        return size
+
+    return get_size_with_aspect_ratio(image_size, size, max_size)
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_numpy_to_framework_fn
+def get_numpy_to_framework_fn(arr) -> Callable:
+    """
+    Returns a function that converts a numpy array to the framework of the input array.
+
+    Args:
+        arr (`np.ndarray`): The array to convert.
+    """
+    if isinstance(arr, np.ndarray):
+        return np.array
+    if is_tf_available() and is_tf_tensor(arr):
+        import tensorflow as tf
+
+        return tf.convert_to_tensor
+    if is_torch_available() and is_torch_tensor(arr):
+        import torch
+
+        return torch.tensor
+    if is_flax_available() and is_jax_tensor(arr):
+        import jax.numpy as jnp
+
+        return jnp.array
+    raise ValueError(f"Cannot convert arrays of type {type(arr)}")
+
+
+# Copied from transformers.models.detr.image_processing_detr.safe_squeeze
+def safe_squeeze(arr: np.ndarray, axis: Optional[int] = None) -> np.ndarray:
+    """
+    Squeezes an array, but only if the axis specified has dim 1.
+    """
+    if axis is None:
+        return arr.squeeze()
+
+    try:
+        return arr.squeeze(axis=axis)
+    except ValueError:
+        return arr
+
+
+# Copied from transformers.models.detr.image_processing_detr.normalize_annotation
+def normalize_annotation(annotation: Dict, image_size: Tuple[int, int]) -> Dict:
+    image_height, image_width = image_size
+    norm_annotation = {}
+    for key, value in annotation.items():
+        if key == "boxes":
+            boxes = value
+            boxes = corners_to_center_format(boxes)
+            boxes /= np.asarray([image_width, image_height, image_width, image_height], dtype=np.float32)
+            norm_annotation[key] = boxes
+        else:
+            norm_annotation[key] = value
+    return norm_annotation
+
+
+# Copied from transformers.models.detr.image_processing_detr.max_across_indices
+def max_across_indices(values: Iterable[Any]) -> List[Any]:
+    """
+    Return the maximum value across all indices of an iterable of values.
+    """
+    return [max(values_i) for values_i in zip(*values)]
+
+
+# Copied from transformers.models.detr.image_processing_detr.get_max_height_width
+def get_max_height_width(images: List[np.ndarray]) -> List[int]:
+    """
+    Get the maximum height and width across all images in a batch.
+    """
+    input_channel_dimension = infer_channel_dimension_format(images[0])
+
+    if input_channel_dimension == ChannelDimension.FIRST:
+        _, max_height, max_width = max_across_indices([img.shape for img in images])
+    elif input_channel_dimension == ChannelDimension.LAST:
+        max_height, max_width, _ = max_across_indices([img.shape for img in images])
+    else:
+        raise ValueError(f"Invalid channel dimension format: {input_channel_dimension}")
+    return (max_height, max_width)
+
+
+# Copied from transformers.models.detr.image_processing_detr.make_pixel_mask
+def make_pixel_mask(image: np.ndarray, output_size: Tuple[int, int]) -> np.ndarray:
+    """
+    Make a pixel mask for the image, where 1 indicates a valid pixel and 0 indicates padding.
+
+    Args:
+        image (`np.ndarray`):
+            Image to make the pixel mask for.
+        output_size (`Tuple[int, int]`):
+            Output size of the mask.
+    """
+    input_height, input_width = get_image_size(image)
+    mask = np.zeros(output_size, dtype=np.int64)
+    mask[:input_height, :input_width] = 1
+    return mask
+
+
+# Copied from transformers.models.detr.image_processing_detr.convert_coco_poly_to_mask
+def convert_coco_poly_to_mask(segmentations, height: int, width: int) -> np.ndarray:
+    """
+    Convert a COCO polygon annotation to a mask.
+
+    Args:
+        segmentations (`List[List[float]]`):
+            List of polygons, each polygon represented by a list of x-y coordinates.
+        height (`int`):
+            Height of the mask.
+        width (`int`):
+            Width of the mask.
+    """
+    try:
+        from pycocotools import mask as coco_mask
+    except ImportError:
+        raise ImportError("Pycocotools is not installed in your environment.")
+
+    masks = []
+    for polygons in segmentations:
+        rles = coco_mask.frPyObjects(polygons, height, width)
+        mask = coco_mask.decode(rles)
+        if len(mask.shape) < 3:
+            mask = mask[..., None]
+        mask = np.asarray(mask, dtype=np.uint8)
+        mask = np.any(mask, axis=2)
+        masks.append(mask)
+    if masks:
+        masks = np.stack(masks, axis=0)
+    else:
+        masks = np.zeros((0, height, width), dtype=np.uint8)
+
+    return masks
+
+
+# Copied from transformers.models.detr.image_processing_detr.prepare_coco_detection_annotation with DETR->DETA
+def prepare_coco_detection_annotation(image, target, return_segmentation_masks: bool = False):
+    """
+    Convert the target in COCO format into the format expected by DETA.
+    """
+    image_height, image_width = get_image_size(image)
+
+    image_id = target["image_id"]
+    image_id = np.asarray([image_id], dtype=np.int64)
+
+    # Get all COCO annotations for the given image.
+    annotations = target["annotations"]
+    annotations = [obj for obj in annotations if "iscrowd" not in obj or obj["iscrowd"] == 0]
+
+    classes = [obj["category_id"] for obj in annotations]
+    classes = np.asarray(classes, dtype=np.int64)
+
+    # for conversion to coco api
+    area = np.asarray([obj["area"] for obj in annotations], dtype=np.float32)
+    iscrowd = np.asarray([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in annotations], dtype=np.int64)
+
+    boxes = [obj["bbox"] for obj in annotations]
+    # guard against no boxes via resizing
+    boxes = np.asarray(boxes, dtype=np.float32).reshape(-1, 4)
+    boxes[:, 2:] += boxes[:, :2]
+    boxes[:, 0::2] = boxes[:, 0::2].clip(min=0, max=image_width)
+    boxes[:, 1::2] = boxes[:, 1::2].clip(min=0, max=image_height)
+
+    keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
+
+    new_target = {}
+    new_target["image_id"] = image_id
+    new_target["class_labels"] = classes[keep]
+    new_target["boxes"] = boxes[keep]
+    new_target["area"] = area[keep]
+    new_target["iscrowd"] = iscrowd[keep]
+    new_target["orig_size"] = np.asarray([int(image_height), int(image_width)], dtype=np.int64)
+
+    if annotations and "keypoints" in annotations[0]:
+        keypoints = [obj["keypoints"] for obj in annotations]
+        keypoints = np.asarray(keypoints, dtype=np.float32)
+        num_keypoints = keypoints.shape[0]
+        keypoints = keypoints.reshape((-1, 3)) if num_keypoints else keypoints
+        new_target["keypoints"] = keypoints[keep]
+
+    if return_segmentation_masks:
+        segmentation_masks = [obj["segmentation"] for obj in annotations]
+        masks = convert_coco_poly_to_mask(segmentation_masks, image_height, image_width)
+        new_target["masks"] = masks[keep]
+
+    return new_target
+
+
+# Copied from transformers.models.detr.image_processing_detr.masks_to_boxes
+def masks_to_boxes(masks: np.ndarray) -> np.ndarray:
+    """
+    Compute the bounding boxes around the provided panoptic segmentation masks.
+
+    Args:
+        masks: masks in format `[number_masks, height, width]` where N is the number of masks
+
+    Returns:
+        boxes: bounding boxes in format `[number_masks, 4]` in xyxy format
+    """
+    if masks.size == 0:
+        return np.zeros((0, 4))
+
+    h, w = masks.shape[-2:]
+    y = np.arange(0, h, dtype=np.float32)
+    x = np.arange(0, w, dtype=np.float32)
+    # see https://github.com/pytorch/pytorch/issues/50276
+    y, x = np.meshgrid(y, x, indexing="ij")
+
+    x_mask = masks * np.expand_dims(x, axis=0)
+    x_max = x_mask.reshape(x_mask.shape[0], -1).max(-1)
+    x = np.ma.array(x_mask, mask=~(np.array(masks, dtype=bool)))
+    x_min = x.filled(fill_value=1e8)
+    x_min = x_min.reshape(x_min.shape[0], -1).min(-1)
+
+    y_mask = masks * np.expand_dims(y, axis=0)
+    y_max = y_mask.reshape(x_mask.shape[0], -1).max(-1)
+    y = np.ma.array(y_mask, mask=~(np.array(masks, dtype=bool)))
+    y_min = y.filled(fill_value=1e8)
+    y_min = y_min.reshape(y_min.shape[0], -1).min(-1)
+
+    return np.stack([x_min, y_min, x_max, y_max], 1)
+
+
+# Copied from transformers.models.detr.image_processing_detr.prepare_coco_panoptic_annotation with DETR->DETA
+def prepare_coco_panoptic_annotation(
+    image: np.ndarray, target: Dict, masks_path: Union[str, pathlib.Path], return_masks: bool = True
+) -> Dict:
+    """
+    Prepare a coco panoptic annotation for DETA.
+    """
+    image_height, image_width = get_image_size(image)
+    annotation_path = pathlib.Path(masks_path) / target["file_name"]
+
+    new_target = {}
+    new_target["image_id"] = np.asarray([target["image_id"] if "image_id" in target else target["id"]], dtype=np.int64)
+    new_target["size"] = np.asarray([image_height, image_width], dtype=np.int64)
+    new_target["orig_size"] = np.asarray([image_height, image_width], dtype=np.int64)
+
+    if "segments_info" in target:
+        masks = np.asarray(PIL.Image.open(annotation_path), dtype=np.uint32)
+        masks = rgb_to_id(masks)
+
+        ids = np.array([segment_info["id"] for segment_info in target["segments_info"]])
+        masks = masks == ids[:, None, None]
+        masks = masks.astype(np.uint8)
+        if return_masks:
+            new_target["masks"] = masks
+        new_target["boxes"] = masks_to_boxes(masks)
+        new_target["class_labels"] = np.array(
+            [segment_info["category_id"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["iscrowd"] = np.asarray(
+            [segment_info["iscrowd"] for segment_info in target["segments_info"]], dtype=np.int64
+        )
+        new_target["area"] = np.asarray(
+            [segment_info["area"] for segment_info in target["segments_info"]], dtype=np.float32
+        )
+
+    return new_target
+
+
+# Copied from transformers.models.detr.image_processing_detr.resize_annotation
+def resize_annotation(
+    annotation: Dict[str, Any],
+    orig_size: Tuple[int, int],
+    target_size: Tuple[int, int],
+    threshold: float = 0.5,
+    resample: PILImageResampling = PILImageResampling.NEAREST,
+):
+    """
+    Resizes an annotation to a target size.
+
+    Args:
+        annotation (`Dict[str, Any]`):
+            The annotation dictionary.
+        orig_size (`Tuple[int, int]`):
+            The original size of the input image.
+        target_size (`Tuple[int, int]`):
+            The target size of the image, as returned by the preprocessing `resize` step.
+        threshold (`float`, *optional*, defaults to 0.5):
+            The threshold used to binarize the segmentation masks.
+        resample (`PILImageResampling`, defaults to `PILImageResampling.NEAREST`):
+            The resampling filter to use when resizing the masks.
+    """
+    ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(target_size, orig_size))
+    ratio_height, ratio_width = ratios
+
+    new_annotation = {}
+    new_annotation["size"] = target_size
+
+    for key, value in annotation.items():
+        if key == "boxes":
+            boxes = value
+            scaled_boxes = boxes * np.asarray([ratio_width, ratio_height, ratio_width, ratio_height], dtype=np.float32)
+            new_annotation["boxes"] = scaled_boxes
+        elif key == "area":
+            area = value
+            scaled_area = area * (ratio_width * ratio_height)
+            new_annotation["area"] = scaled_area
+        elif key == "masks":
+            masks = value[:, None]
+            masks = np.array([resize(mask, target_size, resample=resample) for mask in masks])
+            masks = masks.astype(np.float32)
+            masks = masks[:, 0] > threshold
+            new_annotation["masks"] = masks
+        elif key == "size":
+            new_annotation["size"] = target_size
+        else:
+            new_annotation[key] = value
+
+    return new_annotation
+
+
+class DetaImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a Deformable DETR image processor.
+
+    Args:
+        format (`str`, *optional*, defaults to `"coco_detection"`):
+            Data format of the annotations. One of "coco_detection" or "coco_panoptic".
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Controls whether to resize the image's (height, width) dimensions to the specified `size`. Can be
+            overridden by the `do_resize` parameter in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"shortest_edge": 800, "longest_edge": 1333}`):
+            Size of the image's (height, width) dimensions after resizing. Can be overridden by the `size` parameter in
+            the `preprocess` method.
+        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BILINEAR`):
+            Resampling filter to use if resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Controls whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by the
+            `do_rescale` parameter in the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by the `rescale_factor` parameter in the
+            `preprocess` method.
+        do_normalize:
+            Controls whether to normalize the image. Can be overridden by the `do_normalize` parameter in the
+            `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_DEFAULT_MEAN`):
+            Mean values to use when normalizing the image. Can be a single value or a list of values, one for each
+            channel. Can be overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_DEFAULT_STD`):
+            Standard deviation values to use when normalizing the image. Can be a single value or a list of values, one
+            for each channel. Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_pad (`bool`, *optional*, defaults to `True`):
+            Controls whether to pad the image to the largest image in a batch and create a pixel mask. Can be
+            overridden by the `do_pad` parameter in the `preprocess` method.
+    """
+
+    model_input_names = ["pixel_values", "pixel_mask"]
+
+    def __init__(
+        self,
+        format: Union[str, AnnotionFormat] = AnnotionFormat.COCO_DETECTION,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Union[float, List[float]] = None,
+        image_std: Union[float, List[float]] = None,
+        do_pad: bool = True,
+        **kwargs
+    ) -> None:
+        if "pad_and_return_pixel_mask" in kwargs:
+            do_pad = kwargs.pop("pad_and_return_pixel_mask")
+
+        size = size if size is not None else {"shortest_edge": 800, "longest_edge": 1333}
+        size = get_size_dict(size, default_to_square=False)
+
+        super().__init__(**kwargs)
+        self.format = format
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else IMAGENET_DEFAULT_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_DEFAULT_STD
+        self.do_pad = do_pad
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.prepare_annotation with DETR->DETA
+    def prepare_annotation(
+        self,
+        image: np.ndarray,
+        target: Dict,
+        format: Optional[AnnotionFormat] = None,
+        return_segmentation_masks: bool = None,
+        masks_path: Optional[Union[str, pathlib.Path]] = None,
+    ) -> Dict:
+        """
+        Prepare an annotation for feeding into DETA model.
+        """
+        format = format if format is not None else self.format
+
+        if format == AnnotionFormat.COCO_DETECTION:
+            return_segmentation_masks = False if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_detection_annotation(image, target, return_segmentation_masks)
+        elif format == AnnotionFormat.COCO_PANOPTIC:
+            return_segmentation_masks = True if return_segmentation_masks is None else return_segmentation_masks
+            target = prepare_coco_panoptic_annotation(
+                image, target, masks_path=masks_path, return_masks=return_segmentation_masks
+            )
+        else:
+            raise ValueError(f"Format {format} is not supported.")
+        return target
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.prepare
+    def prepare(self, image, target, return_segmentation_masks=None, masks_path=None):
+        warnings.warn(
+            "The `prepare` method is deprecated and will be removed in a future version. "
+            "Please use `prepare_annotation` instead. Note: the `prepare_annotation` method "
+            "does not return the image anymore.",
+        )
+        target = self.prepare_annotation(image, target, return_segmentation_masks, masks_path, self.format)
+        return image, target
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.convert_coco_poly_to_mask
+    def convert_coco_poly_to_mask(self, *args, **kwargs):
+        warnings.warn("The `convert_coco_poly_to_mask` method is deprecated and will be removed in a future version. ")
+        return convert_coco_poly_to_mask(*args, **kwargs)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.prepare_coco_detection
+    def prepare_coco_detection(self, *args, **kwargs):
+        warnings.warn("The `prepare_coco_detection` method is deprecated and will be removed in a future version. ")
+        return prepare_coco_detection_annotation(*args, **kwargs)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.prepare_coco_panoptic
+    def prepare_coco_panoptic(self, *args, **kwargs):
+        warnings.warn("The `prepare_coco_panoptic` method is deprecated and will be removed in a future version. ")
+        return prepare_coco_panoptic_annotation(*args, **kwargs)
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BILINEAR,
+        data_format: Optional[ChannelDimension] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Resize the image to the given size. Size can be `min_size` (scalar) or `(height, width)` tuple. If size is an
+        int, smaller edge of the image will be matched to this number.
+        """
+        size = get_size_dict(size, default_to_square=False)
+        if "shortest_edge" in size and "longest_edge" in size:
+            size = get_resize_output_image_size(image, size["shortest_edge"], size["longest_edge"])
+        elif "height" in size and "width" in size:
+            size = (size["height"], size["width"])
+        else:
+            raise ValueError(
+                "Size must contain 'height' and 'width' keys or 'shortest_edge' and 'longest_edge' keys. Got"
+                f" {size.keys()}."
+            )
+        image = resize(image, size=size, resample=resample, data_format=data_format)
+        return image
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.resize_annotation
+    def resize_annotation(
+        self,
+        annotation,
+        orig_size,
+        size,
+        resample: PILImageResampling = PILImageResampling.NEAREST,
+    ) -> Dict:
+        """
+        Resize the annotation to match the resized image. If size is an int, smaller edge of the mask will be matched
+        to this number.
+        """
+        return resize_annotation(annotation, orig_size=orig_size, target_size=size, resample=resample)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.rescale
+    def rescale(
+        self, image: np.ndarray, rescale_factor: Union[float, int], data_format: Optional[ChannelDimension] = None
+    ) -> np.ndarray:
+        """
+        Rescale the image by the given factor.
+        """
+        return rescale(image, rescale_factor, data_format=data_format)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.normalize
+    def normalize(
+        self,
+        image: np.ndarray,
+        mean: Union[float, Iterable[float]],
+        std: Union[float, Iterable[float]],
+        data_format: Optional[ChannelDimension] = None,
+    ) -> np.ndarray:
+        """
+        Normalize the image with the given mean and standard deviation.
+        """
+        return normalize(image, mean=mean, std=std, data_format=data_format)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.normalize_annotation
+    def normalize_annotation(self, annotation: Dict, image_size: Tuple[int, int]) -> Dict:
+        """
+        Normalize the boxes in the annotation from `[top_left_x, top_left_y, bottom_right_x, bottom_right_y]` to
+        `[center_x, center_y, width, height]` format.
+        """
+        return normalize_annotation(annotation, image_size=image_size)
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.pad_and_create_pixel_mask
+    def pad_and_create_pixel_mask(
+        self,
+        pixel_values_list: List[ImageInput],
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = None,
+    ) -> BatchFeature:
+        """
+        Pads a batch of images with zeros to the size of largest height and width in the batch and returns their
+        corresponding pixel mask.
+
+        Args:
+            images (`List[np.ndarray]`):
+                Batch of images to pad.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                    - Unset: Return a list of `np.ndarray`.
+                    - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                    - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                    - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                    - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        warnings.warn(
+            "This method is deprecated and will be removed in v4.27.0. Please use pad instead.", FutureWarning
+        )
+        # pad expects a list of np.ndarray, but the previous feature extractors expected torch tensors
+        images = [to_numpy_array(image) for image in pixel_values_list]
+        return self.pad(
+            images=images,
+            return_pixel_mask=True,
+            return_tensors=return_tensors,
+            data_format=data_format,
+        )
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor._pad_image
+    def _pad_image(
+        self,
+        image: np.ndarray,
+        output_size: Tuple[int, int],
+        constant_values: Union[float, Iterable[float]] = 0,
+        data_format: Optional[ChannelDimension] = None,
+    ) -> np.ndarray:
+        """
+        Pad an image with zeros to the given size.
+        """
+        input_height, input_width = get_image_size(image)
+        output_height, output_width = output_size
+
+        pad_bottom = output_height - input_height
+        pad_right = output_width - input_width
+        padding = ((0, pad_bottom), (0, pad_right))
+        padded_image = pad(
+            image, padding, mode=PaddingMode.CONSTANT, constant_values=constant_values, data_format=data_format
+        )
+        return padded_image
+
+    # Copied from transformers.models.detr.image_processing_detr.DetrImageProcessor.pad
+    def pad(
+        self,
+        images: List[np.ndarray],
+        constant_values: Union[float, Iterable[float]] = 0,
+        return_pixel_mask: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = None,
+    ) -> np.ndarray:
+        """
+        Pads a batch of images to the bottom and right of the image with zeros to the size of largest height and width
+        in the batch and optionally returns their corresponding pixel mask.
+
+        Args:
+            image (`np.ndarray`):
+                Image to pad.
+            constant_values (`float` or `Iterable[float]`, *optional*):
+                The value to use for the padding if `mode` is `"constant"`.
+            return_pixel_mask (`bool`, *optional*, defaults to `True`):
+                Whether to return a pixel mask.
+            input_channel_dimension (`ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be inferred from the input image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        pad_size = get_max_height_width(images)
+
+        padded_images = [
+            self._pad_image(image, pad_size, constant_values=constant_values, data_format=data_format)
+            for image in images
+        ]
+        data = {"pixel_values": padded_images}
+
+        if return_pixel_mask:
+            masks = [make_pixel_mask(image=image, output_size=pad_size) for image in images]
+            data["pixel_mask"] = masks
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        annotations: Optional[Union[List[Dict], List[List[Dict]]]] = None,
+        return_segmentation_masks: bool = None,
+        masks_path: Optional[Union[str, pathlib.Path]] = None,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample=None,  # PILImageResampling
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[Union[int, float]] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_pad: Optional[bool] = None,
+        format: Optional[Union[str, AnnotionFormat]] = None,
+        return_tensors: Optional[Union[TensorType, str]] = None,
+        data_format: Union[str, ChannelDimension] = ChannelDimension.FIRST,
+        **kwargs
+    ) -> BatchFeature:
+        """
+        Preprocess an image or a batch of images so that it can be used by the model.
+
+        Args:
+            images (`ImageInput`):
+                Image or batch of images to preprocess.
+            annotations (`List[Dict]` or `List[List[Dict]]`, *optional*):
+                List of annotations associated with the image or batch of images. If annotionation is for object
+                detection, the annotations should be a dictionary with the following keys:
+                - "image_id" (`int`): The image id.
+                - "annotations" (`List[Dict]`): List of annotations for an image. Each annotation should be a
+                  dictionary. An image can have no annotations, in which case the list should be empty.
+                If annotionation is for segmentation, the annotations should be a dictionary with the following keys:
+                - "image_id" (`int`): The image id.
+                - "segments_info" (`List[Dict]`): List of segments for an image. Each segment should be a dictionary.
+                  An image can have no segments, in which case the list should be empty.
+                - "file_name" (`str`): The file name of the image.
+            return_segmentation_masks (`bool`, *optional*, defaults to self.return_segmentation_masks):
+                Whether to return segmentation masks.
+            masks_path (`str` or `pathlib.Path`, *optional*):
+                Path to the directory containing the segmentation masks.
+            do_resize (`bool`, *optional*, defaults to self.do_resize):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to self.size):
+                Size of the image after resizing.
+            resample (`PILImageResampling`, *optional*, defaults to self.resample):
+                Resampling filter to use when resizing the image.
+            do_rescale (`bool`, *optional*, defaults to self.do_rescale):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to self.rescale_factor):
+                Rescale factor to use when rescaling the image.
+            do_normalize (`bool`, *optional*, defaults to self.do_normalize):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to self.image_mean):
+                Mean to use when normalizing the image.
+            image_std (`float` or `List[float]`, *optional*, defaults to self.image_std):
+                Standard deviation to use when normalizing the image.
+            do_pad (`bool`, *optional*, defaults to self.do_pad):
+                Whether to pad the image.
+            format (`str` or `AnnotionFormat`, *optional*, defaults to self.format):
+                Format of the annotations.
+            return_tensors (`str` or `TensorType`, *optional*, defaults to self.return_tensors):
+                Type of tensors to return. If `None`, will return the list of images.
+            data_format (`str` or `ChannelDimension`, *optional*, defaults to self.data_format):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        if "pad_and_return_pixel_mask" in kwargs:
+            warnings.warn(
+                "The `pad_and_return_pixel_mask` argument is deprecated and will be removed in a future version, "
+                "use `do_pad` instead.",
+                FutureWarning,
+            )
+            do_pad = kwargs.pop("pad_and_return_pixel_mask")
+
+        do_resize = self.do_resize if do_resize is None else do_resize
+        size = self.size if size is None else size
+        size = get_size_dict(size=size, default_to_square=False)
+        resample = self.resample if resample is None else resample
+        do_rescale = self.do_rescale if do_rescale is None else do_rescale
+        rescale_factor = self.rescale_factor if rescale_factor is None else rescale_factor
+        do_normalize = self.do_normalize if do_normalize is None else do_normalize
+        image_mean = self.image_mean if image_mean is None else image_mean
+        image_std = self.image_std if image_std is None else image_std
+        do_pad = self.do_pad if do_pad is None else do_pad
+        format = self.format if format is None else format
+
+        if do_resize is not None and size is None:
+            raise ValueError("Size and max_size must be specified if do_resize is True.")
+
+        if do_rescale is not None and rescale_factor is None:
+            raise ValueError("Rescale factor must be specified if do_rescale is True.")
+
+        if do_normalize is not None and (image_mean is None or image_std is None):
+            raise ValueError("Image mean and std must be specified if do_normalize is True.")
+
+        if not is_batched(images):
+            images = [images]
+            annotations = [annotations] if annotations is not None else None
+
+        if annotations is not None and len(images) != len(annotations):
+            raise ValueError(
+                f"The number of images ({len(images)}) and annotations ({len(annotations)}) do not match."
+            )
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        format = AnnotionFormat(format)
+        if annotations is not None:
+            if format == AnnotionFormat.COCO_DETECTION and not valid_coco_detection_annotations(annotations):
+                raise ValueError(
+                    "Invalid COCO detection annotations. Annotations must a dict (single image) of list of dicts"
+                    "(batch of images) with the following keys: `image_id` and `annotations`, with the latter "
+                    "being a list of annotations in the COCO format."
+                )
+            elif format == AnnotionFormat.COCO_PANOPTIC and not valid_coco_panoptic_annotations(annotations):
+                raise ValueError(
+                    "Invalid COCO panoptic annotations. Annotations must a dict (single image) of list of dicts "
+                    "(batch of images) with the following keys: `image_id`, `file_name` and `segments_info`, with "
+                    "the latter being a list of annotations in the COCO format."
+                )
+            elif format not in SUPPORTED_ANNOTATION_FORMATS:
+                raise ValueError(
+                    f"Unsupported annotation format: {format} must be one of {SUPPORTED_ANNOTATION_FORMATS}"
+                )
+
+        if (
+            masks_path is not None
+            and format == AnnotionFormat.COCO_PANOPTIC
+            and not isinstance(masks_path, (pathlib.Path, str))
+        ):
+            raise ValueError(
+                "The path to the directory containing the mask PNG files should be provided as a"
+                f" `pathlib.Path` or string object, but is {type(masks_path)} instead."
+            )
+
+        # All transformations expect numpy arrays
+        images = [to_numpy_array(image) for image in images]
+
+        # prepare (COCO annotations as a list of Dict -> DETR target as a single Dict per image)
+        if annotations is not None:
+            prepared_images = []
+            prepared_annotations = []
+            for image, target in zip(images, annotations):
+                target = self.prepare_annotation(
+                    image, target, format, return_segmentation_masks=return_segmentation_masks, masks_path=masks_path
+                )
+                prepared_images.append(image)
+                prepared_annotations.append(target)
+            images = prepared_images
+            annotations = prepared_annotations
+            del prepared_images, prepared_annotations
+
+        # transformations
+        if do_resize:
+            if annotations is not None:
+                resized_images, resized_annotations = [], []
+                for image, target in zip(images, annotations):
+                    orig_size = get_image_size(image)
+                    resized_image = self.resize(image, size=size, resample=resample)
+                    resized_annotation = self.resize_annotation(target, orig_size, get_image_size(resized_image))
+                    resized_images.append(resized_image)
+                    resized_annotations.append(resized_annotation)
+                images = resized_images
+                annotations = resized_annotations
+                del resized_images, resized_annotations
+            else:
+                images = [self.resize(image, size=size, resample=resample) for image in images]
+
+        if do_rescale:
+            images = [self.rescale(image, rescale_factor) for image in images]
+
+        if do_normalize:
+            images = [self.normalize(image, image_mean, image_std) for image in images]
+            if annotations is not None:
+                annotations = [
+                    self.normalize_annotation(annotation, get_image_size(image))
+                    for annotation, image in zip(annotations, images)
+                ]
+
+        if do_pad:
+            # Pads images and returns their mask: {'pixel_values': ..., 'pixel_mask': ...}
+            data = self.pad(images, return_pixel_mask=True, data_format=data_format)
+        else:
+            images = [to_channel_dimension_format(image, data_format) for image in images]
+            data = {"pixel_values": images}
+
+        encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
+        if annotations is not None:
+            encoded_inputs["labels"] = [
+                BatchFeature(annotation, tensor_type=return_tensors) for annotation in annotations
+            ]
+
+        return encoded_inputs
+
+    def post_process_object_detection(
+        self,
+        outputs,
+        threshold: float = 0.5,
+        target_sizes: Union[TensorType, List[Tuple]] = None,
+        nms_threshold: float = 0.7,
+    ):
+        """
+        Converts the output of [`DetaForObjectDetection`] into final bounding boxes in (top_left_x, top_left_y,
+        bottom_right_x, bottom_right_y) format. Only supports PyTorch.
+
+        Args:
+            outputs ([`DetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*, defaults to 0.5):
+                Score threshold to keep object detection predictions.
+            target_sizes (`torch.Tensor` or `List[Tuple[int, int]]`, *optional*):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`Tuple[int, int]`) containing the target size
+                (height, width) of each image in the batch. If left to None, predictions will not be resized.
+            nms_threshold (`float`, *optional*, defaults to 0.7):
+                NMS threshold.
+
+        Returns:
+            `List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+        batch_size, num_queries, num_labels = out_logits.shape
+
+        if target_sizes is not None:
+            if len(out_logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+        prob = out_logits.sigmoid()
+
+        all_scores = prob.view(batch_size, num_queries * num_labels).to(out_logits.device)
+        all_indexes = torch.arange(num_queries * num_labels)[None].repeat(batch_size, 1).to(out_logits.device)
+        all_boxes = all_indexes // out_logits.shape[2]
+        all_labels = all_indexes % out_logits.shape[2]
+
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, all_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            if isinstance(target_sizes, List):
+                img_h = torch.Tensor([i[0] for i in target_sizes])
+                img_w = torch.Tensor([i[1] for i in target_sizes])
+            else:
+                img_h, img_w = target_sizes.unbind(1)
+
+            scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1).to(boxes.device)
+            boxes = boxes * scale_fct[:, None, :]
+
+        results = []
+        for b in range(batch_size):
+            box = boxes[b]
+            score = all_scores[b]
+            lbls = all_labels[b]
+
+            pre_topk = score.topk(min(10000, len(score))).indices
+            box = box[pre_topk]
+            score = score[pre_topk]
+            lbls = lbls[pre_topk]
+
+            # apply NMS
+            keep_inds = batched_nms(box, score, lbls, nms_threshold)[:100]
+            score = score[keep_inds]
+            lbls = lbls[keep_inds]
+            box = box[keep_inds]
+
+            results.append(
+                {
+                    "scores": score[score > threshold],
+                    "labels": lbls[score > threshold],
+                    "boxes": box[score > threshold],
+                }
+            )
+
+        return results

src/transformers/models/deta/modeling_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.
+""" PyTorch DETA model."""
+
+
+import copy
+import math
+import warnings
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+
+from ...activations import ACT2FN
+from ...file_utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_scipy_available,
+    is_vision_available,
+    replace_return_docstrings,
+)
+from ...modeling_outputs import BaseModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import meshgrid
+from ...utils import is_torchvision_available, logging, requires_backends
+from ..auto import AutoBackbone
+from .configuration_deta import DetaConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    from transformers.image_transforms import center_to_corners_format
+
+if is_torchvision_available():
+    from torchvision.ops.boxes import batched_nms
+
+if is_scipy_available():
+    from scipy.optimize import linear_sum_assignment
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "DetaConfig"
+_CHECKPOINT_FOR_DOC = "jozhang97/deta-swin-large-o365"
+
+DETA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "jozhang97/deta-swin-large-o365",
+    # See all DETA models at https://huggingface.co/models?filter=deta
+]
+
+
+@dataclass
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrDecoderOutput with DeformableDetr->Deta
+class DetaDecoderOutput(ModelOutput):
+    """
+    Base class for outputs of the DetaDecoder. This class adds two attributes to BaseModelOutputWithCrossAttentions,
+    namely:
+    - a stacked tensor of intermediate decoder hidden states (i.e. the output of each decoder layer)
+    - a stacked tensor of intermediate reference points.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+            Stacked intermediate hidden states (output of each layer of the decoder).
+        intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, sequence_length, hidden_size)`):
+            Stacked intermediate reference points (reference points of each layer of the decoder).
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer
+            plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
+            the self-attention heads.
+        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax,
+            used to compute the weighted average in the cross-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    intermediate_hidden_states: torch.FloatTensor = None
+    intermediate_reference_points: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModelOutput with DeformableDetr->Deta,Deformable DETR->DETA
+class DetaModelOutput(ModelOutput):
+    """
+    Base class for outputs of the Deformable DETR encoder-decoder model.
+
+    Args:
+        init_reference_points (`torch.FloatTensor` of shape  `(batch_size, num_queries, 4)`):
+            Initial reference points sent through the Transformer decoder.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+        intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+            Stacked intermediate hidden states (output of each layer of the decoder).
+        intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+            Stacked intermediate reference points (reference points of each layer of the decoder).
+        decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, num_queries, hidden_size)`. Hidden-states of the decoder at the output of each layer
+            plus the initial embedding outputs.
+        decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, num_queries,
+            num_queries)`. Attentions weights of the decoder, after the attention softmax, used to compute the weighted
+            average in the self-attention heads.
+        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_queries, num_heads, 4, 4)`.
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each
+            layer plus the initial embedding outputs.
+        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_queries, num_heads, 4, 4)`.
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+            Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+            picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+            foreground and background).
+        enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+            Logits of predicted bounding boxes coordinates in the first stage.
+    """
+
+    init_reference_points: torch.FloatTensor = None
+    last_hidden_state: torch.FloatTensor = None
+    intermediate_hidden_states: torch.FloatTensor = None
+    intermediate_reference_points: torch.FloatTensor = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    enc_outputs_class: Optional[torch.FloatTensor] = None
+    enc_outputs_coord_logits: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrObjectDetectionOutput with DeformableDetr->Deta
+class DetaObjectDetectionOutput(ModelOutput):
+    """
+    Output type of [`DetaForObjectDetection`].
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` are provided)):
+            Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a
+            bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized
+            scale-invariant IoU loss.
+        loss_dict (`Dict`, *optional*):
+            A dictionary containing the individual losses. Useful for logging.
+        logits (`torch.FloatTensor` of shape `(batch_size, num_queries, num_classes + 1)`):
+            Classification logits (including no-object) for all queries.
+        pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
+            Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
+            values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
+            possible padding). You can use [`~DetaProcessor.post_process_object_detection`] to retrieve the
+            unnormalized bounding boxes.
+        auxiliary_outputs (`list[Dict]`, *optional*):
+            Optional, only returned when auxilary losses are activated (i.e. `config.auxiliary_loss` is set to `True`)
+            and labels are provided. It is a list of dictionaries containing the two above keys (`logits` and
+            `pred_boxes`) for each decoder layer.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+        decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, num_queries, hidden_size)`. Hidden-states of the decoder at the output of each layer
+            plus the initial embedding outputs.
+        decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, num_queries,
+            num_queries)`. Attentions weights of the decoder, after the attention softmax, used to compute the weighted
+            average in the self-attention heads.
+        cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_queries, num_heads, 4, 4)`.
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each
+            layer plus the initial embedding outputs.
+        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_heads, 4,
+            4)`. Attentions weights of the encoder, after the attention softmax, used to compute the weighted average
+            in the self-attention heads.
+        intermediate_hidden_states (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, hidden_size)`):
+            Stacked intermediate hidden states (output of each layer of the decoder).
+        intermediate_reference_points (`torch.FloatTensor` of shape `(batch_size, config.decoder_layers, num_queries, 4)`):
+            Stacked intermediate reference points (reference points of each layer of the decoder).
+        init_reference_points (`torch.FloatTensor` of shape  `(batch_size, num_queries, 4)`):
+            Initial reference points sent through the Transformer decoder.
+        enc_outputs_class (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.num_labels)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+            Predicted bounding boxes scores where the top `config.two_stage_num_proposals` scoring bounding boxes are
+            picked as region proposals in the first stage. Output of bounding box binary classification (i.e.
+            foreground and background).
+        enc_outputs_coord_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, 4)`, *optional*, returned when `config.with_box_refine=True` and `config.two_stage=True`):
+            Logits of predicted bounding boxes coordinates in the first stage.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    loss_dict: Optional[Dict] = None
+    logits: torch.FloatTensor = None
+    pred_boxes: torch.FloatTensor = None
+    auxiliary_outputs: Optional[List[Dict]] = None
+    init_reference_points: Optional[torch.FloatTensor] = None
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    intermediate_hidden_states: Optional[torch.FloatTensor] = None
+    intermediate_reference_points: Optional[torch.FloatTensor] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    enc_outputs_class: Optional = None
+    enc_outputs_coord_logits: Optional = None
+
+
+def _get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def inverse_sigmoid(x, eps=1e-5):
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+# Copied from transformers.models.detr.modeling_detr.DetrFrozenBatchNorm2d with Detr->Deta
+class DetaFrozenBatchNorm2d(nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than
+    torchvision.models.resnet[18,34,50,101] produce nans.
+    """
+
+    def __init__(self, n):
+        super().__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        num_batches_tracked_key = prefix + "num_batches_tracked"
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super()._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+        )
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it user-friendly
+        weight = self.weight.reshape(1, -1, 1, 1)
+        bias = self.bias.reshape(1, -1, 1, 1)
+        running_var = self.running_var.reshape(1, -1, 1, 1)
+        running_mean = self.running_mean.reshape(1, -1, 1, 1)
+        epsilon = 1e-5
+        scale = weight * (running_var + epsilon).rsqrt()
+        bias = bias - running_mean * scale
+        return x * scale + bias
+
+
+# Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->Deta
+def replace_batch_norm(m, name=""):
+    for attr_str in dir(m):
+        target_attr = getattr(m, attr_str)
+        if isinstance(target_attr, nn.BatchNorm2d):
+            frozen = DetaFrozenBatchNorm2d(target_attr.num_features)
+            bn = getattr(m, attr_str)
+            frozen.weight.data.copy_(bn.weight)
+            frozen.bias.data.copy_(bn.bias)
+            frozen.running_mean.data.copy_(bn.running_mean)
+            frozen.running_var.data.copy_(bn.running_var)
+            setattr(m, attr_str, frozen)
+    for n, ch in m.named_children():
+        replace_batch_norm(ch, n)
+
+
+class DetaBackboneWithPositionalEncodings(nn.Module):
+    """
+    Backbone model with positional embeddings.
+
+    nn.BatchNorm2d layers are replaced by DetaFrozenBatchNorm2d as defined above.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        backbone = AutoBackbone.from_config(config.backbone_config)
+        with torch.no_grad():
+            replace_batch_norm(backbone)
+        self.model = backbone
+        self.intermediate_channel_sizes = self.model.channels
+
+        # TODO fix this
+        if config.backbone_config.model_type == "resnet":
+            for name, parameter in self.model.named_parameters():
+                if "stages.1" not in name and "stages.2" not in name and "stages.3" not in name:
+                    parameter.requires_grad_(False)
+
+        self.position_embedding = build_position_encoding(config)
+
+    def forward(self, pixel_values: torch.Tensor, pixel_mask: torch.Tensor):
+        """
+        Outputs feature maps of latter stages C_3 through C_5 in ResNet if `config.num_feature_levels > 1`, otherwise
+        outputs feature maps of C_5.
+        """
+        # first, send pixel_values through the backbone to get list of feature maps
+        features = self.model(pixel_values).feature_maps
+
+        # next, create position embeddings
+        out = []
+        pos = []
+        for feature_map in features:
+            # downsample pixel_mask to match shape of corresponding feature_map
+            mask = nn.functional.interpolate(pixel_mask[None].float(), size=feature_map.shape[-2:]).to(torch.bool)[0]
+            position_embeddings = self.position_embedding(feature_map, mask).to(feature_map.dtype)
+            out.append((feature_map, mask))
+            pos.append(position_embeddings)
+
+        return out, pos
+
+
+# Copied from transformers.models.detr.modeling_detr._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, target_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[batch_size, seq_len]` to `[batch_size, 1, target_seq_len, source_seq_len]`.
+    """
+    batch_size, source_len = mask.size()
+    target_len = target_len if target_len is not None else source_len
+
+    expanded_mask = mask[:, None, None, :].expand(batch_size, 1, target_len, source_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrSinePositionEmbedding with DeformableDetr->Deta
+class DetaSinePositionEmbedding(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one used by the Attention is all you
+    need paper, generalized to work on images.
+    """
+
+    def __init__(self, embedding_dim=64, temperature=10000, normalize=False, scale=None):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+    def forward(self, pixel_values, pixel_mask):
+        if pixel_mask is None:
+            raise ValueError("No pixel mask provided")
+        y_embed = pixel_mask.cumsum(1, dtype=torch.float32)
+        x_embed = pixel_mask.cumsum(2, dtype=torch.float32)
+        if self.normalize:
+            eps = 1e-6
+            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.embedding_dim, dtype=torch.float32, device=pixel_values.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.embedding_dim)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        return pos
+
+
+# Copied from transformers.models.detr.modeling_detr.DetrLearnedPositionEmbedding
+class DetaLearnedPositionEmbedding(nn.Module):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, embedding_dim=256):
+        super().__init__()
+        self.row_embeddings = nn.Embedding(50, embedding_dim)
+        self.column_embeddings = nn.Embedding(50, embedding_dim)
+
+    def forward(self, pixel_values, pixel_mask=None):
+        height, width = pixel_values.shape[-2:]
+        width_values = torch.arange(width, device=pixel_values.device)
+        height_values = torch.arange(height, device=pixel_values.device)
+        x_emb = self.column_embeddings(width_values)
+        y_emb = self.row_embeddings(height_values)
+        pos = torch.cat([x_emb.unsqueeze(0).repeat(height, 1, 1), y_emb.unsqueeze(1).repeat(1, width, 1)], dim=-1)
+        pos = pos.permute(2, 0, 1)
+        pos = pos.unsqueeze(0)
+        pos = pos.repeat(pixel_values.shape[0], 1, 1, 1)
+        return pos
+
+
+# Copied from transformers.models.detr.modeling_detr.build_position_encoding with Detr->Deta
+def build_position_encoding(config):
+    n_steps = config.d_model // 2
+    if config.position_embedding_type == "sine":
+        # TODO find a better way of exposing other arguments
+        position_embedding = DetaSinePositionEmbedding(n_steps, normalize=True)
+    elif config.position_embedding_type == "learned":
+        position_embedding = DetaLearnedPositionEmbedding(n_steps)
+    else:
+        raise ValueError(f"Not supported {config.position_embedding_type}")
+
+    return position_embedding
+
+
+# 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
+    _, 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 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[level_id].flatten(2).transpose(1, 2).reshape(batch_size * num_heads, hidden_dim, height, width)
+        )
+        # 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_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_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_heads * hidden_dim, num_queries)
+    )
+    return output.transpose(1, 2).contiguous()
+
+
+class DetaMultiscaleDeformableAttention(nn.Module):
+    """
+    Multiscale deformable attention as proposed in Deformable DETR.
+    """
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrMultiscaleDeformableAttention.__init__ with DeformableDetr->Deta
+    def __init__(self, embed_dim: int, num_heads: int, n_levels: int, n_points: int):
+        super().__init__()
+        if embed_dim % num_heads != 0:
+            raise ValueError(
+                f"embed_dim (d_model) must be divisible by num_heads, but got {embed_dim} and {num_heads}"
+            )
+        dim_per_head = embed_dim // num_heads
+        # check if dim_per_head is power of 2
+        if not ((dim_per_head & (dim_per_head - 1) == 0) and dim_per_head != 0):
+            warnings.warn(
+                "You'd better set embed_dim (d_model) in DetaMultiscaleDeformableAttention to make the"
+                " dimension of each attention head a power of 2 which is more efficient in the authors' CUDA"
+                " implementation."
+            )
+
+        self.im2col_step = 64
+
+        self.d_model = embed_dim
+        self.n_levels = n_levels
+        self.n_heads = num_heads
+        self.n_points = n_points
+
+        self.sampling_offsets = nn.Linear(embed_dim, num_heads * n_levels * n_points * 2)
+        self.attention_weights = nn.Linear(embed_dim, num_heads * n_levels * n_points)
+        self.value_proj = nn.Linear(embed_dim, embed_dim)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        nn.init.constant_(self.sampling_offsets.weight.data, 0.0)
+        thetas = torch.arange(self.n_heads, dtype=torch.float32) * (2.0 * math.pi / self.n_heads)
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (
+            (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
+            .view(self.n_heads, 1, 1, 2)
+            .repeat(1, self.n_levels, self.n_points, 1)
+        )
+        for i in range(self.n_points):
+            grid_init[:, :, i, :] *= i + 1
+        with torch.no_grad():
+            self.sampling_offsets.bias = nn.Parameter(grid_init.view(-1))
+        nn.init.constant_(self.attention_weights.weight.data, 0.0)
+        nn.init.constant_(self.attention_weights.bias.data, 0.0)
+        nn.init.xavier_uniform_(self.value_proj.weight.data)
+        nn.init.constant_(self.value_proj.bias.data, 0.0)
+        nn.init.xavier_uniform_(self.output_proj.weight.data)
+        nn.init.constant_(self.output_proj.bias.data, 0.0)
+
+    def with_pos_embed(self, tensor: torch.Tensor, position_embeddings: Optional[Tensor]):
+        return tensor if position_embeddings is None else tensor + position_embeddings
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        position_embeddings: Optional[torch.Tensor] = None,
+        reference_points=None,
+        spatial_shapes=None,
+        level_start_index=None,
+        output_attentions: bool = False,
+    ):
+        # add position embeddings to the hidden states before projecting to queries and keys
+        if position_embeddings is not None:
+            hidden_states = self.with_pos_embed(hidden_states, position_embeddings)
+
+        batch_size, num_queries, _ = hidden_states.shape
+        batch_size, sequence_length, _ = encoder_hidden_states.shape
+        if (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() != sequence_length:
+            raise ValueError(
+                "Make sure to align the spatial shapes with the sequence length of the encoder hidden states"
+            )
+
+        value = self.value_proj(encoder_hidden_states)
+        if attention_mask is not None:
+            # we invert the attention_mask
+            value = value.masked_fill(~attention_mask[..., None], float(0))
+        value = value.view(batch_size, sequence_length, self.n_heads, self.d_model // self.n_heads)
+        sampling_offsets = self.sampling_offsets(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points, 2
+        )
+        attention_weights = self.attention_weights(hidden_states).view(
+            batch_size, num_queries, self.n_heads, self.n_levels * self.n_points
+        )
+        attention_weights = F.softmax(attention_weights, -1).view(
+            batch_size, num_queries, self.n_heads, self.n_levels, self.n_points
+        )
+        # batch_size, num_queries, n_heads, n_levels, n_points, 2
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack([spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :]
+                + sampling_offsets / offset_normalizer[None, None, None, :, None, :]
+            )
+        elif reference_points.shape[-1] == 4:
+            sampling_locations = (
+                reference_points[:, :, None, :, None, :2]
+                + sampling_offsets / self.n_points * reference_points[:, :, None, :, None, 2:] * 0.5
+            )
+        else:
+            raise ValueError(f"Last dim of reference_points must be 2 or 4, but got {reference_points.shape[-1]}")
+        # PyTorch implementation (for now)
+        output = multi_scale_deformable_attention(value, spatial_shapes, sampling_locations, attention_weights)
+        output = self.output_proj(output)
+
+        return output, attention_weights
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrMultiheadAttention with DeformableDetr->Deta,Deformable DETR->DETA
+class DetaMultiheadAttention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper.
+
+    Here, we add position embeddings to the queries and keys (as explained in the Deformable DETR paper).
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        if self.head_dim * num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, batch_size: int):
+        return tensor.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def with_pos_embed(self, tensor: torch.Tensor, position_embeddings: Optional[Tensor]):
+        return tensor if position_embeddings is None else tensor + position_embeddings
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        batch_size, target_len, embed_dim = hidden_states.size()
+        # add position embeddings to the hidden states before projecting to queries and keys
+        if position_embeddings is not None:
+            hidden_states_original = hidden_states
+            hidden_states = self.with_pos_embed(hidden_states, position_embeddings)
+
+        # get queries, keys and values
+        query_states = self.q_proj(hidden_states) * self.scaling
+        key_states = self._shape(self.k_proj(hidden_states), -1, batch_size)
+        value_states = self._shape(self.v_proj(hidden_states_original), -1, batch_size)
+
+        proj_shape = (batch_size * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, target_len, batch_size).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        source_len = key_states.size(1)
+
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (batch_size * self.num_heads, target_len, source_len):
+            raise ValueError(
+                f"Attention weights should be of size {(batch_size * self.num_heads, target_len, source_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [batch_size, seq_len] -> [batch_size, 1, target_seq_len, source_seq_len]
+            attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (batch_size, 1, target_len, source_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(batch_size, 1, target_len, source_len)}, but is"
+                    f" {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(batch_size, self.num_heads, target_len, source_len) + attention_mask
+            attn_weights = attn_weights.view(batch_size * self.num_heads, target_len, source_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(batch_size, self.num_heads, target_len, source_len)
+            attn_weights = attn_weights_reshaped.view(batch_size * self.num_heads, target_len, source_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (batch_size * self.num_heads, target_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(batch_size, self.num_heads, target_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(batch_size, self.num_heads, target_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(batch_size, target_len, embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrEncoderLayer with DeformableDetr->Deta
+class DetaEncoderLayer(nn.Module):
+    def __init__(self, config: DetaConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = DetaMultiscaleDeformableAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            n_levels=config.num_feature_levels,
+            n_points=config.encoder_n_points,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        position_embeddings: torch.Tensor = None,
+        reference_points=None,
+        spatial_shapes=None,
+        level_start_index=None,
+        output_attentions: bool = False,
+    ):
+        """
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Input to the layer.
+            attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
+                Attention mask.
+            position_embeddings (`torch.FloatTensor`, *optional*):
+                Position embeddings, to be added to `hidden_states`.
+            reference_points (`torch.FloatTensor`, *optional*):
+                Reference points.
+            spatial_shapes (`torch.LongTensor`, *optional*):
+                Spatial shapes of the backbone feature maps.
+            level_start_index (`torch.LongTensor`, *optional*):
+                Level start index.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        # Apply Multi-scale Deformable Attention Module on the multi-scale feature maps.
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            position_embeddings=position_embeddings,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            level_start_index=level_start_index,
+            output_attentions=output_attentions,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        if self.training:
+            if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrDecoderLayer with DeformableDetr->Deta
+class DetaDecoderLayer(nn.Module):
+    def __init__(self, config: DetaConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        # self-attention
+        self.self_attn = DetaMultiheadAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        # cross-attention
+        self.encoder_attn = DetaMultiscaleDeformableAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            n_levels=config.num_feature_levels,
+            n_points=config.decoder_n_points,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        # feedforward neural networks
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[torch.Tensor] = None,
+        reference_points=None,
+        spatial_shapes=None,
+        level_start_index=None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ):
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(seq_len, batch, embed_dim)`.
+            position_embeddings (`torch.FloatTensor`, *optional*):
+                Position embeddings that are added to the queries and keys in the self-attention layer.
+            reference_points (`torch.FloatTensor`, *optional*):
+                Reference points.
+            spatial_shapes (`torch.LongTensor`, *optional*):
+                Spatial shapes.
+            level_start_index (`torch.LongTensor`, *optional*):
+                Level start index.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, target_len, source_len)` where padding elements are indicated by very large negative
+                values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            output_attentions=output_attentions,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        second_residual = hidden_states
+
+        # Cross-Attention
+        cross_attn_weights = None
+        hidden_states, cross_attn_weights = self.encoder_attn(
+            hidden_states=hidden_states,
+            attention_mask=encoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            position_embeddings=position_embeddings,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            level_start_index=level_start_index,
+            output_attentions=output_attentions,
+        )
+
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = second_residual + hidden_states
+
+        hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        return outputs
+
+
+# Copied from transformers.models.detr.modeling_detr.DetrClassificationHead
+class DetaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, input_dim: int, inner_dim: int, num_classes: int, pooler_dropout: float):
+        super().__init__()
+        self.dense = nn.Linear(input_dim, inner_dim)
+        self.dropout = nn.Dropout(p=pooler_dropout)
+        self.out_proj = nn.Linear(inner_dim, num_classes)
+
+    def forward(self, hidden_states: torch.Tensor):
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.dense(hidden_states)
+        hidden_states = torch.tanh(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.out_proj(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrPreTrainedModel with DeformableDetr->Deta
+class DetaPreTrainedModel(PreTrainedModel):
+    config_class = DetaConfig
+    base_model_prefix = "model"
+    main_input_name = "pixel_values"
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+
+        if isinstance(module, DetaLearnedPositionEmbedding):
+            nn.init.uniform_(module.row_embeddings.weight)
+            nn.init.uniform_(module.column_embeddings.weight)
+        elif isinstance(module, DetaMultiscaleDeformableAttention):
+            module._reset_parameters()
+        elif isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        if hasattr(module, "reference_points") and not self.config.two_stage:
+            nn.init.xavier_uniform_(module.reference_points.weight.data, gain=1.0)
+            nn.init.constant_(module.reference_points.bias.data, 0.0)
+        if hasattr(module, "level_embed"):
+            nn.init.normal_(module.level_embed)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, DetaDecoder):
+            module.gradient_checkpointing = value
+
+
+DETA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`DetaConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+DETA_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it.
+
+            Pixel values can be obtained using [`AutoImageProcessor`]. See [`AutoImageProcessor.__call__`] for details.
+
+        pixel_mask (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*):
+            Mask to avoid performing attention on padding pixel values. Mask values selected in `[0, 1]`:
+
+            - 1 for pixels that are real (i.e. **not masked**),
+            - 0 for pixels that are padding (i.e. **masked**).
+
+            [What are attention masks?](../glossary#attention-mask)
+
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, num_queries)`, *optional*):
+            Not used by default. Can be used to mask object queries.
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
+            can choose to directly pass a flattened representation of an image.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+            Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
+            embedded representation.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrEncoder with DeformableDetr->Deta
+class DetaEncoder(DetaPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* deformable attention layers. Each layer is a
+    [`DetaEncoderLayer`].
+
+    The encoder updates the flattened multi-scale feature maps through multiple deformable attention layers.
+
+    Args:
+        config: DetaConfig
+    """
+
+    def __init__(self, config: DetaConfig):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layers = nn.ModuleList([DetaEncoderLayer(config) for _ in range(config.encoder_layers)])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        """
+        Get reference points for each feature map. Used in decoder.
+
+        Args:
+            spatial_shapes (`torch.LongTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of each feature map.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`):
+                Valid ratios of each feature map.
+            device (`torch.device`):
+                Device on which to create the tensors.
+        Returns:
+            `torch.FloatTensor` of shape `(batch_size, num_queries, num_feature_levels, 2)`
+        """
+        reference_points_list = []
+        for level, (height, width) in enumerate(spatial_shapes):
+
+            ref_y, ref_x = meshgrid(
+                torch.linspace(0.5, height - 0.5, height, dtype=torch.float32, device=device),
+                torch.linspace(0.5, width - 0.5, width, dtype=torch.float32, device=device),
+                indexing="ij",
+            )
+            # TODO: valid_ratios could be useless here. check https://github.com/fundamentalvision/Deformable-DETR/issues/36
+            ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, level, 1] * height)
+            ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, level, 0] * width)
+            ref = torch.stack((ref_x, ref_y), -1)
+            reference_points_list.append(ref)
+        reference_points = torch.cat(reference_points_list, 1)
+        reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+        return reference_points
+
+    def forward(
+        self,
+        inputs_embeds=None,
+        attention_mask=None,
+        position_embeddings=None,
+        spatial_shapes=None,
+        level_start_index=None,
+        valid_ratios=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Flattened feature map (output of the backbone + projection layer) that is passed to the encoder.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding pixel features. Mask values selected in `[0, 1]`:
+                - 1 for pixel features that are real (i.e. **not masked**),
+                - 0 for pixel features that are padding (i.e. **masked**).
+                [What are attention masks?](../glossary#attention-mask)
+            position_embeddings (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Position embeddings that are added to the queries and keys in each self-attention layer.
+            spatial_shapes (`torch.LongTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of each feature map.
+            level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`):
+                Starting index of each feature map.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`):
+                Ratio of valid area in each feature level.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        hidden_states = inputs_embeds
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        reference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=inputs_embeds.device)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        for i, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            layer_outputs = encoder_layer(
+                hidden_states,
+                attention_mask,
+                position_embeddings=position_embeddings,
+                reference_points=reference_points,
+                spatial_shapes=spatial_shapes,
+                level_start_index=level_start_index,
+                output_attentions=output_attentions,
+            )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrDecoder with DeformableDetr->Deta,Deformable DETR->DETA
+class DetaDecoder(DetaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`DetaDecoderLayer`].
+
+    The decoder updates the query embeddings through multiple self-attention and cross-attention layers.
+
+    Some tweaks for Deformable DETR:
+
+    - `position_embeddings`, `reference_points`, `spatial_shapes` and `valid_ratios` are added to the forward pass.
+    - it also returns a stack of intermediate outputs and reference points from all decoding layers.
+
+    Args:
+        config: DetaConfig
+    """
+
+    def __init__(self, config: DetaConfig):
+        super().__init__(config)
+
+        self.dropout = config.dropout
+        self.layers = nn.ModuleList([DetaDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.gradient_checkpointing = False
+
+        # hack implementation for iterative bounding box refinement and two-stage Deformable DETR
+        self.bbox_embed = None
+        self.class_embed = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        position_embeddings=None,
+        reference_points=None,
+        spatial_shapes=None,
+        level_start_index=None,
+        valid_ratios=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
+                The query embeddings that are passed into the decoder.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing cross-attention on padding pixel_values of the encoder. Mask values selected
+                in `[0, 1]`:
+                - 1 for pixels that are real (i.e. **not masked**),
+                - 0 for pixels that are padding (i.e. **masked**).
+            position_embeddings (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
+                Position embeddings that are added to the queries and keys in each self-attention layer.
+            reference_points (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)` is `as_two_stage` else `(batch_size, num_queries, 2)` or , *optional*):
+                Reference point in range `[0, 1]`, top-left (0,0), bottom-right (1, 1), including padding area.
+            spatial_shapes (`torch.FloatTensor` of shape `(num_feature_levels, 2)`):
+                Spatial shapes of the feature maps.
+            level_start_index (`torch.LongTensor` of shape `(num_feature_levels)`, *optional*):
+                Indexes for the start of each feature level. In range `[0, sequence_length]`.
+            valid_ratios (`torch.FloatTensor` of shape `(batch_size, num_feature_levels, 2)`, *optional*):
+                Ratio of valid area in each feature level.
+
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if inputs_embeds is not None:
+            hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        intermediate = ()
+        intermediate_reference_points = ()
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if reference_points.shape[-1] == 4:
+                reference_points_input = (
+                    reference_points[:, :, None] * torch.cat([valid_ratios, valid_ratios], -1)[:, None]
+                )
+            else:
+                if reference_points.shape[-1] != 2:
+                    raise ValueError("Reference points' last dimension must be of size 2")
+                reference_points_input = reference_points[:, :, None] * valid_ratios[:, None]
+
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                    encoder_hidden_states=encoder_hidden_states,
+                    reference_points=reference_points_input,
+                    spatial_shapes=spatial_shapes,
+                    level_start_index=level_start_index,
+                    encoder_attention_mask=encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            # hack implementation for iterative bounding box refinement
+            if self.bbox_embed is not None:
+                tmp = self.bbox_embed[idx](hidden_states)
+                if reference_points.shape[-1] == 4:
+                    new_reference_points = tmp + inverse_sigmoid(reference_points)
+                    new_reference_points = new_reference_points.sigmoid()
+                else:
+                    if reference_points.shape[-1] != 2:
+                        raise ValueError(
+                            f"Reference points' last dimension must be of size 2, but is {reference_points.shape[-1]}"
+                        )
+                    new_reference_points = tmp
+                    new_reference_points[..., :2] = tmp[..., :2] + inverse_sigmoid(reference_points)
+                    new_reference_points = new_reference_points.sigmoid()
+                reference_points = new_reference_points.detach()
+
+            intermediate += (hidden_states,)
+            intermediate_reference_points += (reference_points,)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        # Keep batch_size as first dimension
+        intermediate = torch.stack(intermediate, dim=1)
+        intermediate_reference_points = torch.stack(intermediate_reference_points, dim=1)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    intermediate,
+                    intermediate_reference_points,
+                    all_hidden_states,
+                    all_self_attns,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return DetaDecoderOutput(
+            last_hidden_state=hidden_states,
+            intermediate_hidden_states=intermediate,
+            intermediate_reference_points=intermediate_reference_points,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The bare DETA Model (consisting of a backbone and encoder-decoder Transformer) outputting raw hidden-states without
+    any specific head on top.
+    """,
+    DETA_START_DOCSTRING,
+)
+class DetaModel(DetaPreTrainedModel):
+    def __init__(self, config: DetaConfig):
+        super().__init__(config)
+
+        if config.two_stage:
+            requires_backends(self, ["torchvision"])
+
+        # Create backbone with positional encoding
+        self.backbone = DetaBackboneWithPositionalEncodings(config)
+        intermediate_channel_sizes = self.backbone.intermediate_channel_sizes
+
+        # Create input projection layers
+        if config.num_feature_levels > 1:
+            num_backbone_outs = len(intermediate_channel_sizes)
+            input_proj_list = []
+            for _ in range(num_backbone_outs):
+                in_channels = intermediate_channel_sizes[_]
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, config.d_model, kernel_size=1),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                )
+            for _ in range(config.num_feature_levels - num_backbone_outs):
+                input_proj_list.append(
+                    nn.Sequential(
+                        nn.Conv2d(in_channels, config.d_model, kernel_size=3, stride=2, padding=1),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                )
+                in_channels = config.d_model
+            self.input_proj = nn.ModuleList(input_proj_list)
+        else:
+            self.input_proj = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        nn.Conv2d(intermediate_channel_sizes[-1], config.d_model, kernel_size=1),
+                        nn.GroupNorm(32, config.d_model),
+                    )
+                ]
+            )
+
+        if not config.two_stage:
+            self.query_position_embeddings = nn.Embedding(config.num_queries, config.d_model * 2)
+
+        self.encoder = DetaEncoder(config)
+        self.decoder = DetaDecoder(config)
+
+        self.level_embed = nn.Parameter(torch.Tensor(config.num_feature_levels, config.d_model))
+
+        if config.two_stage:
+            self.enc_output = nn.Linear(config.d_model, config.d_model)
+            self.enc_output_norm = nn.LayerNorm(config.d_model)
+            self.pos_trans = nn.Linear(config.d_model * 2, config.d_model * 2)
+            self.pos_trans_norm = nn.LayerNorm(config.d_model * 2)
+            self.pix_trans = nn.Linear(config.d_model, config.d_model)
+            self.pix_trans_norm = nn.LayerNorm(config.d_model)
+        else:
+            self.reference_points = nn.Linear(config.d_model, 2)
+
+        self.assign_first_stage = config.assign_first_stage
+        self.two_stage_num_proposals = config.two_stage_num_proposals
+
+        self.post_init()
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_encoder
+    def get_encoder(self):
+        return self.encoder
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_decoder
+    def get_decoder(self):
+        return self.decoder
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.freeze_backbone
+    def freeze_backbone(self):
+        for name, param in self.backbone.conv_encoder.model.named_parameters():
+            param.requires_grad_(False)
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.unfreeze_backbone
+    def unfreeze_backbone(self):
+        for name, param in self.backbone.conv_encoder.model.named_parameters():
+            param.requires_grad_(True)
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_valid_ratio
+    def get_valid_ratio(self, mask):
+        """Get the valid ratio of all feature maps."""
+
+        _, height, width = mask.shape
+        valid_height = torch.sum(mask[:, :, 0], 1)
+        valid_width = torch.sum(mask[:, 0, :], 1)
+        valid_ratio_heigth = valid_height.float() / height
+        valid_ratio_width = valid_width.float() / width
+        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_heigth], -1)
+        return valid_ratio
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_proposal_pos_embed
+    def get_proposal_pos_embed(self, proposals):
+        """Get the position embedding of the proposals."""
+
+        num_pos_feats = 128
+        temperature = 10000
+        scale = 2 * math.pi
+
+        dim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=proposals.device)
+        dim_t = temperature ** (2 * (dim_t // 2) / num_pos_feats)
+        # batch_size, num_queries, 4
+        proposals = proposals.sigmoid() * scale
+        # batch_size, num_queries, 4, 128
+        pos = proposals[:, :, :, None] / dim_t
+        # batch_size, num_queries, 4, 64, 2 -> batch_size, num_queries, 512
+        pos = torch.stack((pos[:, :, :, 0::2].sin(), pos[:, :, :, 1::2].cos()), dim=4).flatten(2)
+        return pos
+
+    def gen_encoder_output_proposals(self, enc_output, padding_mask, spatial_shapes):
+        """Generate the encoder output proposals from encoded enc_output.
+
+        Args:
+            enc_output (Tensor[batch_size, sequence_length, hidden_size]): Output of the encoder.
+            padding_mask (Tensor[batch_size, sequence_length]): Padding mask for `enc_output`.
+            spatial_shapes (Tensor[num_feature_levels, 2]): Spatial shapes of the feature maps.
+
+        Returns:
+            `tuple(torch.FloatTensor)`: A tuple of feature map and bbox prediction.
+                - object_query (Tensor[batch_size, sequence_length, hidden_size]): Object query features. Later used to
+                  directly predict a bounding box. (without the need of a decoder)
+                - output_proposals (Tensor[batch_size, sequence_length, 4]): Normalized proposals, after an inverse
+                  sigmoid.
+        """
+        batch_size = enc_output.shape[0]
+        proposals = []
+        _cur = 0
+        level_ids = []
+        for level, (height, width) in enumerate(spatial_shapes):
+            mask_flatten_ = padding_mask[:, _cur : (_cur + height * width)].view(batch_size, height, width, 1)
+            valid_height = torch.sum(~mask_flatten_[:, :, 0, 0], 1)
+            valid_width = torch.sum(~mask_flatten_[:, 0, :, 0], 1)
+
+            grid_y, grid_x = meshgrid(
+                torch.linspace(0, height - 1, height, dtype=torch.float32, device=enc_output.device),
+                torch.linspace(0, width - 1, width, dtype=torch.float32, device=enc_output.device),
+                indexing="ij",
+            )
+            grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)
+
+            scale = torch.cat([valid_width.unsqueeze(-1), valid_height.unsqueeze(-1)], 1).view(batch_size, 1, 1, 2)
+            grid = (grid.unsqueeze(0).expand(batch_size, -1, -1, -1) + 0.5) / scale
+            width_heigth = torch.ones_like(grid) * 0.05 * (2.0**level)
+            proposal = torch.cat((grid, width_heigth), -1).view(batch_size, -1, 4)
+            proposals.append(proposal)
+            _cur += height * width
+            level_ids.append(grid.new_ones(height * width, dtype=torch.long) * level)
+        output_proposals = torch.cat(proposals, 1)
+        output_proposals_valid = ((output_proposals > 0.01) & (output_proposals < 0.99)).all(-1, keepdim=True)
+        output_proposals = torch.log(output_proposals / (1 - output_proposals))  # inverse sigmoid
+        output_proposals = output_proposals.masked_fill(padding_mask.unsqueeze(-1), float("inf"))
+        output_proposals = output_proposals.masked_fill(~output_proposals_valid, float("inf"))
+
+        # assign each pixel as an object query
+        object_query = enc_output
+        object_query = object_query.masked_fill(padding_mask.unsqueeze(-1), float(0))
+        object_query = object_query.masked_fill(~output_proposals_valid, float(0))
+        object_query = self.enc_output_norm(self.enc_output(object_query))
+        level_ids = torch.cat(level_ids)
+        return object_query, output_proposals, level_ids
+
+    @add_start_docstrings_to_model_forward(DETA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=DetaModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values,
+        pixel_mask=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, DetaModel
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("jozhang97/deta-swin-large-o365")
+        >>> model = DetaModel.from_pretrained("jozhang97/deta-swin-large-o365", two_stage=False)
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+        >>> list(last_hidden_states.shape)
+        [1, 900, 256]
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        batch_size, num_channels, height, width = pixel_values.shape
+        device = pixel_values.device
+
+        if pixel_mask is None:
+            pixel_mask = torch.ones(((batch_size, height, width)), dtype=torch.long, device=device)
+
+        # Extract multi-scale feature maps of same resolution `config.d_model` (cf Figure 4 in paper)
+        # First, sent pixel_values + pixel_mask through Backbone to obtain the features
+        # which is a list of tuples
+        features, position_embeddings_list = self.backbone(pixel_values, pixel_mask)
+
+        # Then, apply 1x1 convolution to reduce the channel dimension to d_model (256 by default)
+        sources = []
+        masks = []
+        for level, (source, mask) in enumerate(features):
+            sources.append(self.input_proj[level](source))
+            masks.append(mask)
+            if mask is None:
+                raise ValueError("No attention mask was provided")
+
+        # Lowest resolution feature maps are obtained via 3x3 stride 2 convolutions on the final stage
+        if self.config.num_feature_levels > len(sources):
+            _len_sources = len(sources)
+            for level in range(_len_sources, self.config.num_feature_levels):
+                if level == _len_sources:
+                    source = self.input_proj[level](features[-1][0])
+                else:
+                    source = self.input_proj[level](sources[-1])
+                mask = nn.functional.interpolate(pixel_mask[None].float(), size=source.shape[-2:]).to(torch.bool)[0]
+                pos_l = self.backbone.position_embedding(source, mask).to(source.dtype)
+                sources.append(source)
+                masks.append(mask)
+                position_embeddings_list.append(pos_l)
+
+        # Create queries
+        query_embeds = None
+        if not self.config.two_stage:
+            query_embeds = self.query_position_embeddings.weight
+
+        # Prepare encoder inputs (by flattening)
+        spatial_shapes = [(source.shape[2:]) for source in sources]
+        source_flatten = [source.flatten(2).transpose(1, 2) for source in sources]
+        mask_flatten = [mask.flatten(1) for mask in masks]
+
+        lvl_pos_embed_flatten = []
+        for level, pos_embed in enumerate(position_embeddings_list):
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)
+            lvl_pos_embed = pos_embed + self.level_embed[level].view(1, 1, -1)
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+
+        source_flatten = torch.cat(source_flatten, 1)
+        mask_flatten = torch.cat(mask_flatten, 1)
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
+        spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=source_flatten.device)
+        level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
+        valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1)
+        valid_ratios = valid_ratios.float()
+
+        # Fourth, sent source_flatten + mask_flatten + lvl_pos_embed_flatten (backbone + proj layer output) through encoder
+        # Also provide spatial_shapes, level_start_index and valid_ratios
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                inputs_embeds=source_flatten,
+                attention_mask=mask_flatten,
+                position_embeddings=lvl_pos_embed_flatten,
+                spatial_shapes=spatial_shapes,
+                level_start_index=level_start_index,
+                valid_ratios=valid_ratios,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # Fifth, prepare decoder inputs
+        batch_size, _, num_channels = encoder_outputs[0].shape
+        enc_outputs_class = None
+        enc_outputs_coord_logits = None
+        if self.config.two_stage:
+            object_query_embedding, output_proposals, level_ids = self.gen_encoder_output_proposals(
+                encoder_outputs[0], ~mask_flatten, spatial_shapes
+            )
+
+            # hack implementation for two-stage DETA
+            # apply a detection head to each pixel (A.4 in paper)
+            # linear projection for bounding box binary classification (i.e. foreground and background)
+            enc_outputs_class = self.decoder.class_embed[-1](object_query_embedding)
+            # 3-layer FFN to predict bounding boxes coordinates (bbox regression branch)
+            delta_bbox = self.decoder.bbox_embed[-1](object_query_embedding)
+            enc_outputs_coord_logits = delta_bbox + output_proposals
+
+            # only keep top scoring `config.two_stage_num_proposals` proposals
+            topk = self.two_stage_num_proposals
+            proposal_logit = enc_outputs_class[..., 0]
+
+            if self.assign_first_stage:
+                proposal_boxes = center_to_corners_format(enc_outputs_coord_logits.sigmoid().float()).clamp(0, 1)
+                topk_proposals = []
+                for b in range(batch_size):
+                    prop_boxes_b = proposal_boxes[b]
+                    prop_logits_b = proposal_logit[b]
+
+                    # pre-nms per-level topk
+                    pre_nms_topk = 1000
+                    pre_nms_inds = []
+                    for lvl in range(len(spatial_shapes)):
+                        lvl_mask = level_ids == lvl
+                        pre_nms_inds.append(torch.topk(prop_logits_b.sigmoid() * lvl_mask, pre_nms_topk)[1])
+                    pre_nms_inds = torch.cat(pre_nms_inds)
+
+                    # nms on topk indices
+                    post_nms_inds = batched_nms(
+                        prop_boxes_b[pre_nms_inds], prop_logits_b[pre_nms_inds], level_ids[pre_nms_inds], 0.9
+                    )
+                    keep_inds = pre_nms_inds[post_nms_inds]
+
+                    if len(keep_inds) < self.two_stage_num_proposals:
+                        print(
+                            f"[WARNING] nms proposals ({len(keep_inds)}) < {self.two_stage_num_proposals}, running"
+                            " naive topk"
+                        )
+                        keep_inds = torch.topk(proposal_logit[b], topk)[1]
+
+                    # keep top Q/L indices for L levels
+                    q_per_l = topk // len(spatial_shapes)
+                    is_level_ordered = (
+                        level_ids[keep_inds][None]
+                        == torch.arange(len(spatial_shapes), device=level_ids.device)[:, None]
+                    )
+                    keep_inds_mask = is_level_ordered & (is_level_ordered.cumsum(1) <= q_per_l)  # LS
+                    keep_inds_mask = keep_inds_mask.any(0)  # S
+
+                    # pad to Q indices (might let ones filtered from pre-nms sneak by... unlikely because we pick high conf anyways)
+                    if keep_inds_mask.sum() < topk:
+                        num_to_add = topk - keep_inds_mask.sum()
+                        pad_inds = (~keep_inds_mask).nonzero()[:num_to_add]
+                        keep_inds_mask[pad_inds] = True
+
+                    keep_inds_topk = keep_inds[keep_inds_mask]
+                    topk_proposals.append(keep_inds_topk)
+                topk_proposals = torch.stack(topk_proposals)
+            else:
+                topk_proposals = torch.topk(enc_outputs_class[..., 0], topk, dim=1)[1]
+
+            topk_coords_logits = torch.gather(
+                enc_outputs_coord_logits, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4)
+            )
+            topk_coords_logits = topk_coords_logits.detach()
+            reference_points = topk_coords_logits.sigmoid()
+            init_reference_points = reference_points
+            pos_trans_out = self.pos_trans_norm(self.pos_trans(self.get_proposal_pos_embed(topk_coords_logits)))
+            query_embed, target = torch.split(pos_trans_out, num_channels, dim=2)
+        else:
+            query_embed, target = torch.split(query_embeds, num_channels, dim=1)
+            query_embed = query_embed.unsqueeze(0).expand(batch_size, -1, -1)
+            target = target.unsqueeze(0).expand(batch_size, -1, -1)
+            reference_points = self.reference_points(query_embed).sigmoid()
+            init_reference_points = reference_points
+
+        decoder_outputs = self.decoder(
+            inputs_embeds=target,
+            position_embeddings=query_embed,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=mask_flatten,
+            reference_points=reference_points,
+            spatial_shapes=spatial_shapes,
+            level_start_index=level_start_index,
+            valid_ratios=valid_ratios,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            enc_outputs = tuple(value for value in [enc_outputs_class, enc_outputs_coord_logits] if value is not None)
+            tuple_outputs = (init_reference_points,) + decoder_outputs + encoder_outputs + enc_outputs
+
+            return tuple_outputs
+
+        return DetaModelOutput(
+            init_reference_points=init_reference_points,
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            intermediate_hidden_states=decoder_outputs.intermediate_hidden_states,
+            intermediate_reference_points=decoder_outputs.intermediate_reference_points,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+            enc_outputs_class=enc_outputs_class,
+            enc_outputs_coord_logits=enc_outputs_coord_logits,
+        )
+
+
+@add_start_docstrings(
+    """
+    DETA Model (consisting of a backbone and encoder-decoder Transformer) with object detection heads on top, for tasks
+    such as COCO detection.
+    """,
+    DETA_START_DOCSTRING,
+)
+class DetaForObjectDetection(DetaPreTrainedModel):
+    # When using clones, all layers > 0 will be clones, but layer 0 *is* required
+    _keys_to_ignore_on_load_missing = ["bbox_embed\.[1-9]\d*", "class_embed\.[1-9]\d*"]
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrForObjectDetection.__init__ with DeformableDetr->Deta
+    def __init__(self, config: DetaConfig):
+        super().__init__(config)
+
+        # Deformable DETR encoder-decoder model
+        self.model = DetaModel(config)
+
+        # Detection heads on top
+        self.class_embed = nn.Linear(config.d_model, config.num_labels)
+        self.bbox_embed = DetaMLPPredictionHead(
+            input_dim=config.d_model, hidden_dim=config.d_model, output_dim=4, num_layers=3
+        )
+
+        prior_prob = 0.01
+        bias_value = -math.log((1 - prior_prob) / prior_prob)
+        self.class_embed.bias.data = torch.ones(config.num_labels) * bias_value
+        nn.init.constant_(self.bbox_embed.layers[-1].weight.data, 0)
+        nn.init.constant_(self.bbox_embed.layers[-1].bias.data, 0)
+
+        # if two-stage, the last class_embed and bbox_embed is for region proposal generation
+        num_pred = (config.decoder_layers + 1) if config.two_stage else config.decoder_layers
+        if config.with_box_refine:
+            self.class_embed = _get_clones(self.class_embed, num_pred)
+            self.bbox_embed = _get_clones(self.bbox_embed, num_pred)
+            nn.init.constant_(self.bbox_embed[0].layers[-1].bias.data[2:], -2.0)
+            # hack implementation for iterative bounding box refinement
+            self.model.decoder.bbox_embed = self.bbox_embed
+        else:
+            nn.init.constant_(self.bbox_embed.layers[-1].bias.data[2:], -2.0)
+            self.class_embed = nn.ModuleList([self.class_embed for _ in range(num_pred)])
+            self.bbox_embed = nn.ModuleList([self.bbox_embed for _ in range(num_pred)])
+            self.model.decoder.bbox_embed = None
+        if config.two_stage:
+            # hack implementation for two-stage
+            self.model.decoder.class_embed = self.class_embed
+            for box_embed in self.bbox_embed:
+                nn.init.constant_(box_embed.layers[-1].bias.data[2:], 0.0)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @torch.jit.unused
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrForObjectDetection._set_aux_loss
+    def _set_aux_loss(self, outputs_class, outputs_coord):
+        # this is a workaround to make torchscript happy, as torchscript
+        # doesn't support dictionary with non-homogeneous values, such
+        # as a dict having both a Tensor and a list.
+        return [{"logits": a, "pred_boxes": b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1])]
+
+    @add_start_docstrings_to_model_forward(DETA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=DetaObjectDetectionOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values,
+        pixel_mask=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (`List[Dict]` of len `(batch_size,)`, *optional*):
+            Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the
+            following 2 keys: 'class_labels' and 'boxes' (the class labels and bounding boxes of an image in the batch
+            respectively). The class labels themselves should be a `torch.LongTensor` of len `(number of bounding boxes
+            in the image,)` and the boxes a `torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)`.
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, DetaForObjectDetection
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> image_processor = AutoImageProcessor.from_pretrained("jozhang97/deta-swin-large")
+        >>> model = DetaForObjectDetection.from_pretrained("jozhang97/deta-swin-large")
+
+        >>> inputs = image_processor(images=image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> # convert outputs (bounding boxes and class logits) to COCO API
+        >>> target_sizes = torch.tensor([image.size[::-1]])
+        >>> results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[
+        ...     0
+        ... ]
+        >>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
+        ...     box = [round(i, 2) for i in box.tolist()]
+        ...     print(
+        ...         f"Detected {model.config.id2label[label.item()]} with confidence "
+        ...         f"{round(score.item(), 3)} at location {box}"
+        ...     )
+        Detected cat with confidence 0.683 at location [345.85, 23.68, 639.86, 372.83]
+        Detected cat with confidence 0.683 at location [8.8, 52.49, 316.93, 473.45]
+        Detected remote with confidence 0.568 at location [40.02, 73.75, 175.96, 117.33]
+        Detected remote with confidence 0.546 at location [333.68, 77.13, 370.12, 187.51]
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # First, sent images through DETR base model to obtain encoder + decoder outputs
+        outputs = self.model(
+            pixel_values,
+            pixel_mask=pixel_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            encoder_outputs=encoder_outputs,
+            inputs_embeds=inputs_embeds,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs.intermediate_hidden_states if return_dict else outputs[2]
+        init_reference = outputs.init_reference_points if return_dict else outputs[0]
+        inter_references = outputs.intermediate_reference_points if return_dict else outputs[3]
+
+        # class logits + predicted bounding boxes
+        outputs_classes = []
+        outputs_coords = []
+
+        for level in range(hidden_states.shape[1]):
+            if level == 0:
+                reference = init_reference
+            else:
+                reference = inter_references[:, level - 1]
+            reference = inverse_sigmoid(reference)
+            outputs_class = self.class_embed[level](hidden_states[:, level])
+            delta_bbox = self.bbox_embed[level](hidden_states[:, level])
+            if reference.shape[-1] == 4:
+                outputs_coord_logits = delta_bbox + reference
+            elif reference.shape[-1] == 2:
+                delta_bbox[..., :2] += reference
+                outputs_coord_logits = delta_bbox
+            else:
+                raise ValueError(f"reference.shape[-1] should be 4 or 2, but got {reference.shape[-1]}")
+            outputs_coord = outputs_coord_logits.sigmoid()
+            outputs_classes.append(outputs_class)
+            outputs_coords.append(outputs_coord)
+        # Keep batch_size as first dimension
+        outputs_class = torch.stack(outputs_classes, dim=1)
+        outputs_coord = torch.stack(outputs_coords, dim=1)
+
+        logits = outputs_class[:, -1]
+        pred_boxes = outputs_coord[:, -1]
+
+        loss, loss_dict, auxiliary_outputs = None, None, None
+        if labels is not None:
+            # First: create the matcher
+            matcher = DetaHungarianMatcher(
+                class_cost=self.config.class_cost, bbox_cost=self.config.bbox_cost, giou_cost=self.config.giou_cost
+            )
+            # Second: create the criterion
+            losses = ["labels", "boxes", "cardinality"]
+            criterion = DetaLoss(
+                matcher=matcher,
+                num_classes=self.config.num_labels,
+                focal_alpha=self.config.focal_alpha,
+                losses=losses,
+                num_queries=self.config.num_queries,
+            )
+            criterion.to(logits.device)
+            # Third: compute the losses, based on outputs and labels
+            outputs_loss = {}
+            outputs_loss["logits"] = logits
+            outputs_loss["pred_boxes"] = pred_boxes
+            if self.config.auxiliary_loss:
+                intermediate = outputs.intermediate_hidden_states if return_dict else outputs[4]
+                outputs_class = self.class_embed(intermediate)
+                outputs_coord = self.bbox_embed(intermediate).sigmoid()
+                auxiliary_outputs = self._set_aux_loss(outputs_class, outputs_coord)
+                outputs_loss["auxiliary_outputs"] = auxiliary_outputs
+            if self.config.two_stage:
+                enc_outputs_coord = outputs.enc_outputs_coord_logits.sigmoid()
+                outputs["enc_outputs"] = {"pred_logits": outputs.enc_outputs_class, "pred_boxes": enc_outputs_coord}
+
+            loss_dict = criterion(outputs_loss, labels)
+            # Fourth: compute total loss, as a weighted sum of the various losses
+            weight_dict = {"loss_ce": 1, "loss_bbox": self.config.bbox_loss_coefficient}
+            weight_dict["loss_giou"] = self.config.giou_loss_coefficient
+            if self.config.auxiliary_loss:
+                aux_weight_dict = {}
+                for i in range(self.config.decoder_layers - 1):
+                    aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
+                weight_dict.update(aux_weight_dict)
+            loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
+
+        if not return_dict:
+            if auxiliary_outputs is not None:
+                output = (logits, pred_boxes) + auxiliary_outputs + outputs
+            else:
+                output = (logits, pred_boxes) + outputs
+            tuple_outputs = ((loss, loss_dict) + output) if loss is not None else output
+
+            return tuple_outputs
+
+        dict_outputs = DetaObjectDetectionOutput(
+            loss=loss,
+            loss_dict=loss_dict,
+            logits=logits,
+            pred_boxes=pred_boxes,
+            auxiliary_outputs=auxiliary_outputs,
+            last_hidden_state=outputs.last_hidden_state,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+            intermediate_hidden_states=outputs.intermediate_hidden_states,
+            intermediate_reference_points=outputs.intermediate_reference_points,
+            init_reference_points=outputs.init_reference_points,
+            enc_outputs_class=outputs.enc_outputs_class,
+            enc_outputs_coord_logits=outputs.enc_outputs_coord_logits,
+        )
+
+        return dict_outputs
+
+
+# Copied from transformers.models.detr.modeling_detr.dice_loss
+def dice_loss(inputs, targets, num_boxes):
+    """
+    Compute the DICE loss, similar to generalized IOU for masks
+
+    Args:
+        inputs: A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets: A float tensor with the same shape as inputs. Stores the binary
+                 classification label for each element in inputs (0 for the negative class and 1 for the positive
+                 class).
+    """
+    inputs = inputs.sigmoid()
+    inputs = inputs.flatten(1)
+    numerator = 2 * (inputs * targets).sum(1)
+    denominator = inputs.sum(-1) + targets.sum(-1)
+    loss = 1 - (numerator + 1) / (denominator + 1)
+    return loss.sum() / num_boxes
+
+
+# Copied from transformers.models.detr.modeling_detr.sigmoid_focal_loss
+def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2):
+    """
+    Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
+
+    Args:
+        inputs (`torch.FloatTensor` of arbitrary shape):
+            The predictions for each example.
+        targets (`torch.FloatTensor` with the same shape as `inputs`)
+            A tensor storing the binary classification label for each element in the `inputs` (0 for the negative class
+            and 1 for the positive class).
+        alpha (`float`, *optional*, defaults to `0.25`):
+            Optional weighting factor in the range (0,1) to balance positive vs. negative examples.
+        gamma (`int`, *optional*, defaults to `2`):
+            Exponent of the modulating factor (1 - p_t) to balance easy vs hard examples.
+
+    Returns:
+        Loss tensor
+    """
+    prob = inputs.sigmoid()
+    ce_loss = nn.functional.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
+    # add modulating factor
+    p_t = prob * targets + (1 - prob) * (1 - targets)
+    loss = ce_loss * ((1 - p_t) ** gamma)
+
+    if alpha >= 0:
+        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
+        loss = alpha_t * loss
+
+    return loss.mean(1).sum() / num_boxes
+
+
+class DetaLoss(nn.Module):
+    """
+    This class computes the losses for `DetaForObjectDetection`. The process happens in two steps: 1) we compute
+    hungarian assignment between ground truth boxes and the outputs of the model 2) we supervise each pair of matched
+    ground-truth / prediction (supervised class and box).
+
+    Args:
+        matcher (`DetaHungarianMatcher`):
+            Module able to compute a matching between targets and proposals.
+        num_classes (`int`):
+            Number of object categories, omitting the special no-object category.
+        focal_alpha (`float`):
+            Alpha parameter in focal loss.
+        losses (`List[str]`):
+            List of all the losses to be applied. See `get_loss` for a list of all available losses.
+    """
+
+    def __init__(
+        self,
+        matcher,
+        num_classes,
+        focal_alpha,
+        losses,
+        num_queries,
+        assign_first_stage=False,
+        assign_second_stage=False,
+    ):
+        super().__init__()
+        self.matcher = matcher
+        self.num_classes = num_classes
+        self.focal_alpha = focal_alpha
+        self.losses = losses
+        self.assign_first_stage = assign_first_stage
+        self.assign_second_stage = assign_second_stage
+
+        if self.assign_first_stage:
+            self.stg1_assigner = DetaStage1Assigner()
+        if self.assign_second_stage:
+            self.stg2_assigner = DetaStage2Assigner(num_queries)
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss.loss_labels
+    def loss_labels(self, outputs, targets, indices, num_boxes):
+        """
+        Classification loss (Binary focal loss) targets dicts must contain the key "class_labels" containing a tensor
+        of dim [nb_target_boxes]
+        """
+        if "logits" not in outputs:
+            raise KeyError("No logits were found in the outputs")
+        source_logits = outputs["logits"]
+
+        idx = self._get_source_permutation_idx(indices)
+        target_classes_o = torch.cat([t["class_labels"][J] for t, (_, J) in zip(targets, indices)])
+        target_classes = torch.full(
+            source_logits.shape[:2], self.num_classes, dtype=torch.int64, device=source_logits.device
+        )
+        target_classes[idx] = target_classes_o
+
+        target_classes_onehot = torch.zeros(
+            [source_logits.shape[0], source_logits.shape[1], source_logits.shape[2] + 1],
+            dtype=source_logits.dtype,
+            layout=source_logits.layout,
+            device=source_logits.device,
+        )
+        target_classes_onehot.scatter_(2, target_classes.unsqueeze(-1), 1)
+
+        target_classes_onehot = target_classes_onehot[:, :, :-1]
+        loss_ce = (
+            sigmoid_focal_loss(source_logits, target_classes_onehot, num_boxes, alpha=self.focal_alpha, gamma=2)
+            * source_logits.shape[1]
+        )
+        losses = {"loss_ce": loss_ce}
+
+        return losses
+
+    @torch.no_grad()
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss.loss_cardinality
+    def loss_cardinality(self, outputs, targets, indices, num_boxes):
+        """
+        Compute the cardinality error, i.e. the absolute error in the number of predicted non-empty boxes.
+
+        This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients.
+        """
+        logits = outputs["logits"]
+        device = logits.device
+        target_lengths = torch.as_tensor([len(v["class_labels"]) for v in targets], device=device)
+        # Count the number of predictions that are NOT "no-object" (which is the last class)
+        card_pred = (logits.argmax(-1) != logits.shape[-1] - 1).sum(1)
+        card_err = nn.functional.l1_loss(card_pred.float(), target_lengths.float())
+        losses = {"cardinality_error": card_err}
+        return losses
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss.loss_boxes
+    def loss_boxes(self, outputs, targets, indices, num_boxes):
+        """
+        Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss.
+
+        Targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]. The target boxes
+        are expected in format (center_x, center_y, w, h), normalized by the image size.
+        """
+        if "pred_boxes" not in outputs:
+            raise KeyError("No predicted boxes found in outputs")
+        idx = self._get_source_permutation_idx(indices)
+        source_boxes = outputs["pred_boxes"][idx]
+        target_boxes = torch.cat([t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0)
+
+        loss_bbox = nn.functional.l1_loss(source_boxes, target_boxes, reduction="none")
+
+        losses = {}
+        losses["loss_bbox"] = loss_bbox.sum() / num_boxes
+
+        loss_giou = 1 - torch.diag(
+            generalized_box_iou(center_to_corners_format(source_boxes), center_to_corners_format(target_boxes))
+        )
+        losses["loss_giou"] = loss_giou.sum() / num_boxes
+        return losses
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss._get_source_permutation_idx
+    def _get_source_permutation_idx(self, indices):
+        # permute predictions following indices
+        batch_idx = torch.cat([torch.full_like(source, i) for i, (source, _) in enumerate(indices)])
+        source_idx = torch.cat([source for (source, _) in indices])
+        return batch_idx, source_idx
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss._get_target_permutation_idx
+    def _get_target_permutation_idx(self, indices):
+        # permute targets following indices
+        batch_idx = torch.cat([torch.full_like(target, i) for i, (_, target) in enumerate(indices)])
+        target_idx = torch.cat([target for (_, target) in indices])
+        return batch_idx, target_idx
+
+    # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrLoss.get_loss
+    def get_loss(self, loss, outputs, targets, indices, num_boxes):
+        loss_map = {
+            "labels": self.loss_labels,
+            "cardinality": self.loss_cardinality,
+            "boxes": self.loss_boxes,
+        }
+        if loss not in loss_map:
+            raise ValueError(f"Loss {loss} not supported")
+        return loss_map[loss](outputs, targets, indices, num_boxes)
+
+    def forward(self, outputs, targets):
+        """
+        This performs the loss computation.
+
+        Args:
+             outputs (`dict`, *optional*):
+                Dictionary of tensors, see the output specification of the model for the format.
+             targets (`List[dict]`, *optional*):
+                List of dicts, such that `len(targets) == batch_size`. The expected keys in each dict depends on the
+                losses applied, see each loss' doc.
+        """
+        outputs_without_aux = {k: v for k, v in outputs.items() if k != "auxiliary_outputs"}
+
+        # Retrieve the matching between the outputs of the last layer and the targets
+        if self.assign_second_stage:
+            indices = self.stg2_assigner(outputs_without_aux, targets)
+        else:
+            indices = self.matcher(outputs_without_aux, targets)
+
+        # Compute the average number of target boxes accross all nodes, for normalization purposes
+        num_boxes = sum(len(t["class_labels"]) for t in targets)
+        num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device)
+        # (Niels): comment out function below, distributed training to be added
+        # if is_dist_avail_and_initialized():
+        #     torch.distributed.all_reduce(num_boxes)
+        # (Niels) in original implementation, num_boxes is divided by get_world_size()
+        num_boxes = torch.clamp(num_boxes, min=1).item()
+
+        # Compute all the requested losses
+        losses = {}
+        for loss in self.losses:
+            losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes))
+
+        # In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
+        if "auxiliary_outputs" in outputs:
+            for i, auxiliary_outputs in enumerate(outputs["auxiliary_outputs"]):
+                if not self.assign_second_stage:
+                    indices = self.matcher(auxiliary_outputs, targets)
+                for loss in self.losses:
+                    l_dict = self.get_loss(loss, auxiliary_outputs, targets, indices, num_boxes)
+                    l_dict = {k + f"_{i}": v for k, v in l_dict.items()}
+                    losses.update(l_dict)
+
+        if "enc_outputs" in outputs:
+            enc_outputs = outputs["enc_outputs"]
+            bin_targets = copy.deepcopy(targets)
+            for bt in bin_targets:
+                bt["labels"] = torch.zeros_like(bt["labels"])
+            if self.assign_first_stage:
+                indices = self.stg1_assigner(enc_outputs, bin_targets)
+            else:
+                indices = self.matcher(enc_outputs, bin_targets)
+            for loss in self.losses:
+                kwargs = {}
+                if loss == "labels":
+                    # Logging is enabled only for the last layer
+                    kwargs["log"] = False
+                l_dict = self.get_loss(loss, enc_outputs, bin_targets, indices, num_boxes, **kwargs)
+                l_dict = {k + "_enc": v for k, v in l_dict.items()}
+                losses.update(l_dict)
+
+        return losses
+
+
+# Copied from transformers.models.detr.modeling_detr.DetrMLPPredictionHead
+class DetaMLPPredictionHead(nn.Module):
+    """
+    Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates,
+    height and width of a bounding box w.r.t. an image.
+
+    Copied from https://github.com/facebookresearch/detr/blob/master/models/detr.py
+
+    """
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x
+
+
+# Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrHungarianMatcher with DeformableDetr->Deta
+class DetaHungarianMatcher(nn.Module):
+    """
+    This class computes an assignment between the targets and the predictions of the network.
+
+    For efficiency reasons, the targets don't include the no_object. Because of this, in general, there are more
+    predictions than targets. In this case, we do a 1-to-1 matching of the best predictions, while the others are
+    un-matched (and thus treated as non-objects).
+
+    Args:
+        class_cost:
+            The relative weight of the classification error in the matching cost.
+        bbox_cost:
+            The relative weight of the L1 error of the bounding box coordinates in the matching cost.
+        giou_cost:
+            The relative weight of the giou loss of the bounding box in the matching cost.
+    """
+
+    def __init__(self, class_cost: float = 1, bbox_cost: float = 1, giou_cost: float = 1):
+        super().__init__()
+        requires_backends(self, ["scipy"])
+
+        self.class_cost = class_cost
+        self.bbox_cost = bbox_cost
+        self.giou_cost = giou_cost
+        if class_cost == 0 and bbox_cost == 0 and giou_cost == 0:
+            raise ValueError("All costs of the Matcher can't be 0")
+
+    @torch.no_grad()
+    def forward(self, outputs, targets):
+        """
+        Args:
+            outputs (`dict`):
+                A dictionary that contains at least these entries:
+                * "logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
+                * "pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates.
+            targets (`List[dict]`):
+                A list of targets (len(targets) = batch_size), where each target is a dict containing:
+                * "class_labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of
+                  ground-truth
+                 objects in the target) containing the class labels
+                * "boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates.
+
+        Returns:
+            `List[Tuple]`: A list of size `batch_size`, containing tuples of (index_i, index_j) where:
+            - index_i is the indices of the selected predictions (in order)
+            - index_j is the indices of the corresponding selected targets (in order)
+            For each batch element, it holds: len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
+        """
+        batch_size, num_queries = outputs["logits"].shape[:2]
+
+        # We flatten to compute the cost matrices in a batch
+        out_prob = outputs["logits"].flatten(0, 1).sigmoid()  # [batch_size * num_queries, num_classes]
+        out_bbox = outputs["pred_boxes"].flatten(0, 1)  # [batch_size * num_queries, 4]
+
+        # Also concat the target labels and boxes
+        target_ids = torch.cat([v["class_labels"] for v in targets])
+        target_bbox = torch.cat([v["boxes"] for v in targets])
+
+        # Compute the classification cost.
+        alpha = 0.25
+        gamma = 2.0
+        neg_cost_class = (1 - alpha) * (out_prob**gamma) * (-(1 - out_prob + 1e-8).log())
+        pos_cost_class = alpha * ((1 - out_prob) ** gamma) * (-(out_prob + 1e-8).log())
+        class_cost = pos_cost_class[:, target_ids] - neg_cost_class[:, target_ids]
+
+        # Compute the L1 cost between boxes
+        bbox_cost = torch.cdist(out_bbox, target_bbox, p=1)
+
+        # Compute the giou cost between boxes
+        giou_cost = -generalized_box_iou(center_to_corners_format(out_bbox), center_to_corners_format(target_bbox))
+
+        # Final cost matrix
+        cost_matrix = self.bbox_cost * bbox_cost + self.class_cost * class_cost + self.giou_cost * giou_cost
+        cost_matrix = cost_matrix.view(batch_size, num_queries, -1).cpu()
+
+        sizes = [len(v["boxes"]) for v in targets]
+        indices = [linear_sum_assignment(c[i]) for i, c in enumerate(cost_matrix.split(sizes, -1))]
+        return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
+
+
+# Copied from transformers.models.detr.modeling_detr._upcast
+def _upcast(t: Tensor) -> Tensor:
+    # Protects from numerical overflows in multiplications by upcasting to the equivalent higher type
+    if t.is_floating_point():
+        return t if t.dtype in (torch.float32, torch.float64) else t.float()
+    else:
+        return t if t.dtype in (torch.int32, torch.int64) else t.int()
+
+
+# Copied from transformers.models.detr.modeling_detr.box_area
+def box_area(boxes: Tensor) -> Tensor:
+    """
+    Computes the area of a set of bounding boxes, which are specified by its (x1, y1, x2, y2) coordinates.
+
+    Args:
+        boxes (`torch.FloatTensor` of shape `(number_of_boxes, 4)`):
+            Boxes for which the area will be computed. They are expected to be in (x1, y1, x2, y2) format with `0 <= x1
+            < x2` and `0 <= y1 < y2`.
+
+    Returns:
+        `torch.FloatTensor`: a tensor containing the area for each box.
+    """
+    boxes = _upcast(boxes)
+    return (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
+
+
+# Copied from transformers.models.detr.modeling_detr.box_iou
+def box_iou(boxes1, boxes2):
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    left_top = torch.max(boxes1[:, None, :2], boxes2[:, :2])  # [N,M,2]
+    right_bottom = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])  # [N,M,2]
+
+    width_height = (right_bottom - left_top).clamp(min=0)  # [N,M,2]
+    inter = width_height[:, :, 0] * width_height[:, :, 1]  # [N,M]
+
+    union = area1[:, None] + area2 - inter
+
+    iou = inter / union
+    return iou, union
+
+
+# Copied from transformers.models.detr.modeling_detr.generalized_box_iou
+def generalized_box_iou(boxes1, boxes2):
+    """
+    Generalized IoU from https://giou.stanford.edu/. The boxes should be in [x0, y0, x1, y1] (corner) format.
+
+    Returns:
+        `torch.FloatTensor`: a [N, M] pairwise matrix, where N = len(boxes1) and M = len(boxes2)
+    """
+    # degenerate boxes gives inf / nan results
+    # so do an early check
+    if not (boxes1[:, 2:] >= boxes1[:, :2]).all():
+        raise ValueError(f"boxes1 must be in [x0, y0, x1, y1] (corner) format, but got {boxes1}")
+    if not (boxes2[:, 2:] >= boxes2[:, :2]).all():
+        raise ValueError(f"boxes2 must be in [x0, y0, x1, y1] (corner) format, but got {boxes2}")
+    iou, union = box_iou(boxes1, boxes2)
+
+    top_left = torch.min(boxes1[:, None, :2], boxes2[:, :2])
+    bottom_right = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
+
+    width_height = (bottom_right - top_left).clamp(min=0)  # [N,M,2]
+    area = width_height[:, :, 0] * width_height[:, :, 1]
+
+    return iou - (area - union) / area
+
+
+# from https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/layers/wrappers.py#L100
+def nonzero_tuple(x):
+    """
+    A 'as_tuple=True' version of torch.nonzero to support torchscript. because of
+    https://github.com/pytorch/pytorch/issues/38718
+    """
+    if torch.jit.is_scripting():
+        if x.dim() == 0:
+            return x.unsqueeze(0).nonzero().unbind(1)
+        return x.nonzero().unbind(1)
+    else:
+        return x.nonzero(as_tuple=True)
+
+
+# from https://github.com/facebookresearch/detectron2/blob/9921a2caa585d4fa66c4b534b6fab6e74d89b582/detectron2/modeling/matcher.py#L9
+class DetaMatcher(object):
+    """
+    This class assigns to each predicted "element" (e.g., a box) a ground-truth element. Each predicted element will
+    have exactly zero or one matches; each ground-truth element may be matched to zero or more predicted elements.
+
+    The matching is determined by the MxN match_quality_matrix, that characterizes how well each (ground-truth,
+    prediction)-pair match each other. For example, if the elements are boxes, this matrix may contain box
+    intersection-over-union overlap values.
+
+    The matcher returns (a) a vector of length N containing the index of the ground-truth element m in [0, M) that
+    matches to prediction n in [0, N). (b) a vector of length N containing the labels for each prediction.
+    """
+
+    def __init__(self, thresholds: List[float], labels: List[int], allow_low_quality_matches: bool = False):
+        """
+        Args:
+            thresholds (`list[float]`):
+                A list of thresholds used to stratify predictions into levels.
+            labels (`list[int`):
+                A list of values to label predictions belonging at each level. A label can be one of {-1, 0, 1}
+                signifying {ignore, negative class, positive class}, respectively.
+            allow_low_quality_matches (`bool`, *optional*, defaults to `False`):
+                If `True`, produce additional matches for predictions with maximum match quality lower than
+                high_threshold. See `set_low_quality_matches_` for more details.
+
+            For example,
+                thresholds = [0.3, 0.5] labels = [0, -1, 1] All predictions with iou < 0.3 will be marked with 0 and
+                thus will be considered as false positives while training. All predictions with 0.3 <= iou < 0.5 will
+                be marked with -1 and thus will be ignored. All predictions with 0.5 <= iou will be marked with 1 and
+                thus will be considered as true positives.
+        """
+        # Add -inf and +inf to first and last position in thresholds
+        thresholds = thresholds[:]
+        if thresholds[0] < 0:
+            raise ValueError("Thresholds should be positive")
+        thresholds.insert(0, -float("inf"))
+        thresholds.append(float("inf"))
+        # Currently torchscript does not support all + generator
+        if not all([low <= high for (low, high) in zip(thresholds[:-1], thresholds[1:])]):
+            raise ValueError("Thresholds should be sorted.")
+        if not all([l in [-1, 0, 1] for l in labels]):
+            raise ValueError("All labels should be either -1, 0 or 1")
+        if len(labels) != len(thresholds) - 1:
+            raise ValueError("Number of labels should be equal to number of thresholds - 1")
+        self.thresholds = thresholds
+        self.labels = labels
+        self.allow_low_quality_matches = allow_low_quality_matches
+
+    def __call__(self, match_quality_matrix):
+        """
+        Args:
+            match_quality_matrix (Tensor[float]): an MxN tensor, containing the
+                pairwise quality between M ground-truth elements and N predicted elements. All elements must be >= 0
+                (due to the us of `torch.nonzero` for selecting indices in `set_low_quality_matches_`).
+
+        Returns:
+            matches (Tensor[int64]): a vector of length N, where matches[i] is a matched
+                ground-truth index in [0, M)
+            match_labels (Tensor[int8]): a vector of length N, where pred_labels[i] indicates
+                whether a prediction is a true or false positive or ignored
+        """
+        assert match_quality_matrix.dim() == 2
+        if match_quality_matrix.numel() == 0:
+            default_matches = match_quality_matrix.new_full((match_quality_matrix.size(1),), 0, dtype=torch.int64)
+            # When no gt boxes exist, we define IOU = 0 and therefore set labels
+            # to `self.labels[0]`, which usually defaults to background class 0
+            # To choose to ignore instead, can make labels=[-1,0,-1,1] + set appropriate thresholds
+            default_match_labels = match_quality_matrix.new_full(
+                (match_quality_matrix.size(1),), self.labels[0], dtype=torch.int8
+            )
+            return default_matches, default_match_labels
+
+        assert torch.all(match_quality_matrix >= 0)
+
+        # match_quality_matrix is M (gt) x N (predicted)
+        # Max over gt elements (dim 0) to find best gt candidate for each prediction
+        matched_vals, matches = match_quality_matrix.max(dim=0)
+
+        match_labels = matches.new_full(matches.size(), 1, dtype=torch.int8)
+
+        for l, low, high in zip(self.labels, self.thresholds[:-1], self.thresholds[1:]):
+            low_high = (matched_vals >= low) & (matched_vals < high)
+            match_labels[low_high] = l
+
+        if self.allow_low_quality_matches:
+            self.set_low_quality_matches_(match_labels, match_quality_matrix)
+
+        return matches, match_labels
+
+    def set_low_quality_matches_(self, match_labels, match_quality_matrix):
+        """
+        Produce additional matches for predictions that have only low-quality matches. Specifically, for each
+        ground-truth G find the set of predictions that have maximum overlap with it (including ties); for each
+        prediction in that set, if it is unmatched, then match it to the ground-truth G.
+
+        This function implements the RPN assignment case (i) in Sec. 3.1.2 of :paper:`Faster R-CNN`.
+        """
+        # For each gt, find the prediction with which it has highest quality
+        highest_quality_foreach_gt, _ = match_quality_matrix.max(dim=1)
+        # Find the highest quality match available, even if it is low, including ties.
+        # Note that the matches qualities must be positive due to the use of
+        # `torch.nonzero`.
+        _, pred_inds_with_highest_quality = nonzero_tuple(match_quality_matrix == highest_quality_foreach_gt[:, None])
+        # If an anchor was labeled positive only due to a low-quality match
+        # with gt_A, but it has larger overlap with gt_B, it's matched index will still be gt_B.
+        # This follows the implementation in Detectron, and is found to have no significant impact.
+        match_labels[pred_inds_with_highest_quality] = 1
+
+
+# from https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/modeling/sampling.py#L9
+def subsample_labels(labels: torch.Tensor, num_samples: int, positive_fraction: float, bg_label: int):
+    """
+    Return `num_samples` (or fewer, if not enough found) random samples from `labels` which is a mixture of positives &
+    negatives. It will try to return as many positives as possible without exceeding `positive_fraction * num_samples`,
+    and then try to fill the remaining slots with negatives.
+
+    Args:
+        labels (Tensor): (N, ) label vector with values:
+            * -1: ignore
+            * bg_label: background ("negative") class
+            * otherwise: one or more foreground ("positive") classes
+        num_samples (int): The total number of labels with value >= 0 to return.
+            Values that are not sampled will be filled with -1 (ignore).
+        positive_fraction (float): The number of subsampled labels with values > 0
+            is `min(num_positives, int(positive_fraction * num_samples))`. The number of negatives sampled is
+            `min(num_negatives, num_samples - num_positives_sampled)`. In order words, if there are not enough
+            positives, the sample is filled with negatives. If there are also not enough negatives, then as many
+            elements are sampled as is possible.
+        bg_label (int): label index of background ("negative") class.
+
+    Returns:
+        pos_idx, neg_idx (Tensor):
+            1D vector of indices. The total length of both is `num_samples` or fewer.
+    """
+    positive = nonzero_tuple((labels != -1) & (labels != bg_label))[0]
+    negative = nonzero_tuple(labels == bg_label)[0]
+
+    num_pos = int(num_samples * positive_fraction)
+    # protect against not enough positive examples
+    num_pos = min(positive.numel(), num_pos)
+    num_neg = num_samples - num_pos
+    # protect against not enough negative examples
+    num_neg = min(negative.numel(), num_neg)
+
+    # randomly select positive and negative examples
+    perm1 = torch.randperm(positive.numel(), device=positive.device)[:num_pos]
+    perm2 = torch.randperm(negative.numel(), device=negative.device)[:num_neg]
+
+    pos_idx = positive[perm1]
+    neg_idx = negative[perm2]
+    return pos_idx, neg_idx
+
+
+def sample_topk_per_gt(pr_inds, gt_inds, iou, k):
+    if len(gt_inds) == 0:
+        return pr_inds, gt_inds
+    # find topk matches for each gt
+    gt_inds2, counts = gt_inds.unique(return_counts=True)
+    scores, pr_inds2 = iou[gt_inds2].topk(k, dim=1)
+    gt_inds2 = gt_inds2[:, None].repeat(1, k)
+
+    # filter to as many matches that gt has
+    pr_inds3 = torch.cat([pr[:c] for c, pr in zip(counts, pr_inds2)])
+    gt_inds3 = torch.cat([gt[:c] for c, gt in zip(counts, gt_inds2)])
+    return pr_inds3, gt_inds3
+
+
+# modified from https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/modeling/roi_heads/roi_heads.py#L123
+class DetaStage2Assigner(nn.Module):
+    def __init__(self, num_queries, max_k=4):
+        super().__init__()
+        self.positive_fraction = 0.25
+        self.bg_label = 400  # number > 91 to filter out later
+        self.batch_size_per_image = num_queries
+        self.proposal_matcher = DetaMatcher(thresholds=[0.6], labels=[0, 1], allow_low_quality_matches=True)
+        self.k = max_k
+
+    def _sample_proposals(self, matched_idxs: torch.Tensor, matched_labels: torch.Tensor, gt_classes: torch.Tensor):
+        """
+        Based on the matching between N proposals and M groundtruth, sample the proposals and set their classification
+        labels.
+
+        Args:
+            matched_idxs (Tensor): a vector of length N, each is the best-matched
+                gt index in [0, M) for each proposal.
+            matched_labels (Tensor): a vector of length N, the matcher's label
+                (one of cfg.MODEL.ROI_HEADS.IOU_LABELS) for each proposal.
+            gt_classes (Tensor): a vector of length M.
+
+        Returns:
+            Tensor: a vector of indices of sampled proposals. Each is in [0, N). Tensor: a vector of the same length,
+            the classification label for
+                each sampled proposal. Each sample is labeled as either a category in [0, num_classes) or the
+                background (num_classes).
+        """
+        has_gt = gt_classes.numel() > 0
+        # Get the corresponding GT for each proposal
+        if has_gt:
+            gt_classes = gt_classes[matched_idxs]
+            # Label unmatched proposals (0 label from matcher) as background (label=num_classes)
+            gt_classes[matched_labels == 0] = self.bg_label
+            # Label ignore proposals (-1 label)
+            gt_classes[matched_labels == -1] = -1
+        else:
+            gt_classes = torch.zeros_like(matched_idxs) + self.bg_label
+
+        sampled_fg_idxs, sampled_bg_idxs = subsample_labels(
+            gt_classes, self.batch_size_per_image, self.positive_fraction, self.bg_label
+        )
+
+        sampled_idxs = torch.cat([sampled_fg_idxs, sampled_bg_idxs], dim=0)
+        return sampled_idxs, gt_classes[sampled_idxs]
+
+    def forward(self, outputs, targets, return_cost_matrix=False):
+        # COCO categories are from 1 to 90. They set num_classes=91 and apply sigmoid.
+
+        bs = len(targets)
+        indices = []
+        ious = []
+        for b in range(bs):
+            iou, _ = box_iou(
+                center_to_corners_format(targets[b]["boxes"]),
+                center_to_corners_format(outputs["init_reference"][b].detach()),
+            )
+            matched_idxs, matched_labels = self.proposal_matcher(
+                iou
+            )  # proposal_id -> highest_iou_gt_id, proposal_id -> [1 if iou > 0.6, 0 ow]
+            (
+                sampled_idxs,
+                sampled_gt_classes,
+            ) = self._sample_proposals(  # list of sampled proposal_ids, sampled_id -> [0, num_classes)+[bg_label]
+                matched_idxs, matched_labels, targets[b]["labels"]
+            )
+            pos_pr_inds = sampled_idxs[sampled_gt_classes != self.bg_label]
+            pos_gt_inds = matched_idxs[pos_pr_inds]
+            pos_pr_inds, pos_gt_inds = self.postprocess_indices(pos_pr_inds, pos_gt_inds, iou)
+            indices.append((pos_pr_inds, pos_gt_inds))
+            ious.append(iou)
+        if return_cost_matrix:
+            return indices, ious
+        return indices
+
+    def postprocess_indices(self, pr_inds, gt_inds, iou):
+        return sample_topk_per_gt(pr_inds, gt_inds, iou, self.k)
+
+
+# modified from https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/modeling/proposal_generator/rpn.py#L181
+class DetaStage1Assigner(nn.Module):
+    def __init__(self, t_low=0.3, t_high=0.7, max_k=4):
+        super().__init__()
+        self.positive_fraction = 0.5
+        self.batch_size_per_image = 256
+        self.k = max_k
+        self.t_low = t_low
+        self.t_high = t_high
+        self.anchor_matcher = DetaMatcher(
+            thresholds=[t_low, t_high], labels=[0, -1, 1], allow_low_quality_matches=True
+        )
+
+    def _subsample_labels(self, label):
+        """
+        Randomly sample a subset of positive and negative examples, and overwrite the label vector to the ignore value
+        (-1) for all elements that are not included in the sample.
+
+        Args:
+            labels (Tensor): a vector of -1, 0, 1. Will be modified in-place and returned.
+        """
+        pos_idx, neg_idx = subsample_labels(label, self.batch_size_per_image, self.positive_fraction, 0)
+        # Fill with the ignore label (-1), then set positive and negative labels
+        label.fill_(-1)
+        label.scatter_(0, pos_idx, 1)
+        label.scatter_(0, neg_idx, 0)
+        return label
+
+    def forward(self, outputs, targets):
+        bs = len(targets)
+        indices = []
+        for b in range(bs):
+            anchors = outputs["anchors"][b]
+            if len(targets[b]["boxes"]) == 0:
+                indices.append(
+                    (
+                        torch.tensor([], dtype=torch.long, device=anchors.device),
+                        torch.tensor([], dtype=torch.long, device=anchors.device),
+                    )
+                )
+                continue
+            iou, _ = box_iou(
+                center_to_corners_format(targets[b]["boxes"]),
+                center_to_corners_format(anchors),
+            )
+            matched_idxs, matched_labels = self.anchor_matcher(
+                iou
+            )  # proposal_id -> highest_iou_gt_id, proposal_id -> [1 if iou > 0.7, 0 if iou < 0.3, -1 ow]
+            matched_labels = self._subsample_labels(matched_labels)
+
+            all_pr_inds = torch.arange(len(anchors))
+            pos_pr_inds = all_pr_inds[matched_labels == 1]
+            pos_gt_inds = matched_idxs[pos_pr_inds]
+            pos_pr_inds, pos_gt_inds = self.postprocess_indices(pos_pr_inds, pos_gt_inds, iou)
+            pos_pr_inds, pos_gt_inds = pos_pr_inds.to(anchors.device), pos_gt_inds.to(anchors.device)
+            indices.append((pos_pr_inds, pos_gt_inds))
+        return indices
+
+    def postprocess_indices(self, pr_inds, gt_inds, iou):
+        return sample_topk_per_gt(pr_inds, gt_inds, iou, self.k)

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))

tests/models/deta/test_modeling_deta.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch DETA model. """
+
+
+import inspect
+import math
+import unittest
+
+from transformers import DetaConfig, is_torch_available, is_torchvision_available, is_vision_available
+from transformers.file_utils import cached_property
+from transformers.testing_utils import require_torchvision, require_vision, slow, torch_device
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
+
+
+if is_torch_available():
+    import torch
+
+if is_torchvision_available():
+    from transformers import DetaForObjectDetection, DetaModel
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoImageProcessor
+
+
+class DetaModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=8,
+        is_training=True,
+        use_labels=True,
+        hidden_size=256,
+        num_hidden_layers=2,
+        num_attention_heads=8,
+        intermediate_size=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        num_queries=12,
+        num_channels=3,
+        image_size=196,
+        n_targets=8,
+        num_labels=91,
+        num_feature_levels=4,
+        encoder_n_points=2,
+        decoder_n_points=6,
+        two_stage=False,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.num_queries = num_queries
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.n_targets = n_targets
+        self.num_labels = num_labels
+        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
+
+        # we also set the expected seq length for both encoder and decoder
+        self.encoder_seq_length = (
+            math.ceil(self.image_size / 8) ** 2
+            + math.ceil(self.image_size / 16) ** 2
+            + math.ceil(self.image_size / 32) ** 2
+            + math.ceil(self.image_size / 64) ** 2
+        )
+        self.decoder_seq_length = self.num_queries
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        pixel_mask = torch.ones([self.batch_size, self.image_size, self.image_size], device=torch_device)
+
+        labels = None
+        if self.use_labels:
+            # labels is a list of Dict (each Dict being the labels for a given example in the batch)
+            labels = []
+            for i in range(self.batch_size):
+                target = {}
+                target["class_labels"] = torch.randint(
+                    high=self.num_labels, size=(self.n_targets,), device=torch_device
+                )
+                target["boxes"] = torch.rand(self.n_targets, 4, device=torch_device)
+                target["masks"] = torch.rand(self.n_targets, self.image_size, self.image_size, device=torch_device)
+                labels.append(target)
+
+        config = self.get_config()
+        return config, pixel_values, pixel_mask, labels
+
+    def get_config(self):
+        return DetaConfig(
+            d_model=self.hidden_size,
+            encoder_layers=self.num_hidden_layers,
+            decoder_layers=self.num_hidden_layers,
+            encoder_attention_heads=self.num_attention_heads,
+            decoder_attention_heads=self.num_attention_heads,
+            encoder_ffn_dim=self.intermediate_size,
+            decoder_ffn_dim=self.intermediate_size,
+            dropout=self.hidden_dropout_prob,
+            attention_dropout=self.attention_probs_dropout_prob,
+            num_queries=self.num_queries,
+            num_labels=self.num_labels,
+            num_feature_levels=self.num_feature_levels,
+            encoder_n_points=self.encoder_n_points,
+            decoder_n_points=self.decoder_n_points,
+            two_stage=self.two_stage,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values, pixel_mask, labels = self.prepare_config_and_inputs()
+        inputs_dict = {"pixel_values": pixel_values, "pixel_mask": pixel_mask}
+        return config, inputs_dict
+
+    def create_and_check_deta_model(self, config, pixel_values, pixel_mask, labels):
+        model = DetaModel(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
+        result = model(pixel_values)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.num_queries, self.hidden_size))
+
+    def create_and_check_deta_object_detection_head_model(self, config, pixel_values, pixel_mask, labels):
+        model = DetaForObjectDetection(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
+        result = model(pixel_values)
+
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels))
+        self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4))
+
+        result = model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels)
+
+        self.parent.assertEqual(result.loss.shape, ())
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels))
+        self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4))
+
+
+@require_torchvision
+class DetaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (DetaModel, DetaForObjectDetection) if is_torchvision_available() else ()
+    is_encoder_decoder = True
+    test_torchscript = False
+    test_pruning = False
+    test_head_masking = False
+    test_missing_keys = False
+
+    # special case for head models
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+
+        if return_labels:
+            if model_class.__name__ == "DetaForObjectDetection":
+                labels = []
+                for i in range(self.model_tester.batch_size):
+                    target = {}
+                    target["class_labels"] = torch.ones(
+                        size=(self.model_tester.n_targets,), device=torch_device, dtype=torch.long
+                    )
+                    target["boxes"] = torch.ones(
+                        self.model_tester.n_targets, 4, device=torch_device, dtype=torch.float
+                    )
+                    target["masks"] = torch.ones(
+                        self.model_tester.n_targets,
+                        self.model_tester.image_size,
+                        self.model_tester.image_size,
+                        device=torch_device,
+                        dtype=torch.float,
+                    )
+                    labels.append(target)
+                inputs_dict["labels"] = labels
+
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = DetaModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DetaConfig, has_text_modality=False)
+
+    def test_config(self):
+        # we don't test common_properties and arguments_init as these don't apply for DETA
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_with_num_labels()
+        self.config_tester.check_config_can_be_init_without_params()
+
+    def test_deta_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deta_model(*config_and_inputs)
+
+    def test_deta_object_detection_head_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deta_object_detection_head_model(*config_and_inputs)
+
+    @unittest.skip(reason="DETA does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="DETA does not have a get_input_embeddings method")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="DETA is not a generative model")
+    def test_generate_without_input_ids(self):
+        pass
+
+    @unittest.skip(reason="DETA does not use token embeddings")
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    @unittest.skip(reason="Feed forward chunking is not implemented")
+    def test_feed_forward_chunking(self):
+        pass
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_attention_heads,
+                    self.model_tester.num_feature_levels,
+                    self.model_tester.encoder_n_points,
+                ],
+            )
+            out_len = len(outputs)
+
+            correct_outlen = 8
+
+            # loss is at first position
+            if "labels" in inputs_dict:
+                correct_outlen += 1  # loss is added to beginning
+            # Object Detection model returns pred_logits and pred_boxes
+            if model_class.__name__ == "DetaForObjectDetection":
+                correct_outlen += 2
+
+            self.assertEqual(out_len, correct_outlen)
+
+            # decoder attentions
+            decoder_attentions = outputs.decoder_attentions
+            self.assertIsInstance(decoder_attentions, (list, tuple))
+            self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(decoder_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, self.model_tester.num_queries, self.model_tester.num_queries],
+            )
+
+            # cross attentions
+            cross_attentions = outputs.cross_attentions
+            self.assertIsInstance(cross_attentions, (list, tuple))
+            self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(cross_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_attention_heads,
+                    self.model_tester.num_feature_levels,
+                    self.model_tester.decoder_n_points,
+                ],
+            )
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            elif self.is_encoder_decoder:
+                added_hidden_states = 2
+            else:
+                added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions
+
+            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_attention_heads,
+                    self.model_tester.num_feature_levels,
+                    self.model_tester.encoder_n_points,
+                ],
+            )
+
+    # removed retain_grad and grad on decoder_hidden_states, as queries don't require grad
+    def test_retain_grad_hidden_states_attentions(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.output_hidden_states = True
+        config.output_attentions = True
+
+        # no need to test all models as different heads yield the same functionality
+        model_class = self.all_model_classes[0]
+        model = model_class(config)
+        model.to(torch_device)
+
+        inputs = self._prepare_for_class(inputs_dict, model_class)
+
+        outputs = model(**inputs)
+
+        # we take the second output since last_hidden_state is the second item
+        output = outputs[1]
+
+        encoder_hidden_states = outputs.encoder_hidden_states[0]
+        encoder_attentions = outputs.encoder_attentions[0]
+        encoder_hidden_states.retain_grad()
+        encoder_attentions.retain_grad()
+
+        decoder_attentions = outputs.decoder_attentions[0]
+        decoder_attentions.retain_grad()
+
+        cross_attentions = outputs.cross_attentions[0]
+        cross_attentions.retain_grad()
+
+        output.flatten()[0].backward(retain_graph=True)
+
+        self.assertIsNotNone(encoder_hidden_states.grad)
+        self.assertIsNotNone(encoder_attentions.grad)
+        self.assertIsNotNone(decoder_attentions.grad)
+        self.assertIsNotNone(cross_attentions.grad)
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            if model.config.is_encoder_decoder:
+                expected_arg_names = ["pixel_values", "pixel_mask"]
+                expected_arg_names.extend(
+                    ["head_mask", "decoder_head_mask", "encoder_outputs"]
+                    if "head_mask" and "decoder_head_mask" in arg_names
+                    else []
+                )
+                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+            else:
+                expected_arg_names = ["pixel_values", "pixel_mask"]
+                self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    @unittest.skip(reason="Model doesn't use tied weights")
+    def test_tied_model_weights_key_ignore(self):
+        pass
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if param.requires_grad:
+                        if (
+                            "level_embed" in name
+                            or "sampling_offsets.bias" in name
+                            or "value_proj" in name
+                            or "output_proj" in name
+                            or "reference_points" in name
+                        ):
+                            continue
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
+
+TOLERANCE = 1e-4
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@require_torchvision
+@require_vision
+@slow
+class DetaModelIntegrationTests(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return AutoImageProcessor.from_pretrained("jozhang97/deta-resnet-50") if is_vision_available() else None
+
+    def test_inference_object_detection_head(self):
+        model = DetaForObjectDetection.from_pretrained("jozhang97/deta-resnet-50").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        expected_shape_logits = torch.Size((1, 300, model.config.num_labels))
+        self.assertEqual(outputs.logits.shape, expected_shape_logits)
+
+        expected_logits = torch.tensor(
+            [[-7.3978, -2.5406, -4.1668], [-8.2684, -3.9933, -3.8096], [-7.0515, -3.7973, -5.8516]]
+        ).to(torch_device)
+        expected_boxes = torch.tensor(
+            [[0.5043, 0.4973, 0.9998], [0.2542, 0.5489, 0.4748], [0.5490, 0.2765, 0.0570]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_logits, atol=1e-4))
+
+        expected_shape_boxes = torch.Size((1, 300, 4))
+        self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
+        self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
+
+        # verify postprocessing
+        results = image_processor.post_process_object_detection(
+            outputs, threshold=0.3, target_sizes=[image.size[::-1]]
+        )[0]
+        expected_scores = torch.tensor([0.6392, 0.6276, 0.5546, 0.5260, 0.4706], device=torch_device)
+        expected_labels = [75, 17, 17, 75, 63]
+        expected_slice_boxes = torch.tensor([40.5866, 73.2107, 176.1421, 117.1751], device=torch_device)
+
+        self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
+        self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
+        self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))
+
+    def test_inference_object_detection_head_swin_backbone(self):
+        model = DetaForObjectDetection.from_pretrained("jozhang97/deta-swin-large").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        expected_shape_logits = torch.Size((1, 300, model.config.num_labels))
+        self.assertEqual(outputs.logits.shape, expected_shape_logits)
+
+        expected_logits = torch.tensor(
+            [[-7.6308, -2.8485, -5.3737], [-7.2037, -4.5505, -4.8027], [-7.2943, -4.2611, -4.6617]]
+        ).to(torch_device)
+        expected_boxes = torch.tensor(
+            [[0.4987, 0.4969, 0.9999], [0.2549, 0.5498, 0.4805], [0.5498, 0.2757, 0.0569]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_logits, atol=1e-4))
+
+        expected_shape_boxes = torch.Size((1, 300, 4))
+        self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
+        self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
+
+        # verify postprocessing
+        results = image_processor.post_process_object_detection(
+            outputs, threshold=0.3, target_sizes=[image.size[::-1]]
+        )[0]
+        expected_scores = torch.tensor([0.6831, 0.6826, 0.5684, 0.5464, 0.4392], device=torch_device)
+        expected_labels = [17, 17, 75, 75, 63]
+        expected_slice_boxes = torch.tensor([345.8478, 23.6754, 639.8562, 372.8265], device=torch_device)
+
+        self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
+        self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
+        self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))

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