added base code to work from dfine model

docs/source/en/_toctree.yml

@@ -699,6 +699,8 @@
         title: DAB-DETR
       - local: model_doc/deformable_detr
         title: Deformable DETR
+      - local: model_doc/deim
+        title: deim
       - local: model_doc/deit
         title: DeiT
       - local: model_doc/depth_anything

docs/source/en/model_doc/deim.md

@@ -0,0 +1,48 @@
+<!--Copyright 2025 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# deim
+
+## Overview
+
+The deim model was proposed in [<INSERT PAPER NAME HERE>](<INSERT PAPER LINK HERE>) by <INSERT AUTHORS HERE>.
+<INSERT SHORT SUMMARY HERE>
+
+The abstract from the paper is the following:
+
+*<INSERT PAPER ABSTRACT HERE>*
+
+Tips:
+
+<INSERT TIPS ABOUT MODEL HERE>
+
+This model was contributed by [INSERT YOUR HF USERNAME HERE](https://huggingface.co/<INSERT YOUR HF USERNAME HERE>).
+The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>).
+
+
+## DeimConfig
+
+[[autodoc]] DeimConfig
+
+## DeimModel
+
+[[autodoc]] DeimModel
+    - forward
+
+## DeimForObjectDetection
+
+[[autodoc]] DeimForObjectDetection
+    - forward

src/transformers/models/auto/configuration_auto.py

@@ -83,6 +83,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("ctrl", "CTRLConfig"),
         ("cvt", "CvtConfig"),
         ("d_fine", "DFineConfig"),
+        ("deim", "DeimConfig"),
         ("dab-detr", "DabDetrConfig"),
         ("dac", "DacConfig"),
         ("data2vec-audio", "Data2VecAudioConfig"),
@@ -439,6 +440,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("ctrl", "CTRL"),
         ("cvt", "CvT"),
         ("d_fine", "D-FINE"),
+        ("deim", "deim"),
         ("dab-detr", "DAB-DETR"),
         ("dac", "DAC"),
         ("data2vec-audio", "Data2VecAudio"),

src/transformers/models/auto/modeling_auto.py

@@ -81,6 +81,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("ctrl", "CTRLModel"),
         ("cvt", "CvtModel"),
         ("d_fine", "DFineModel"),
+        ("deim", "DeimModel"),
         ("dab-detr", "DabDetrModel"),
         ("dac", "DacModel"),
         ("data2vec-audio", "Data2VecAudioModel"),
@@ -949,6 +950,7 @@ MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
         # Model for Object Detection mapping
         ("conditional_detr", "ConditionalDetrForObjectDetection"),
         ("d_fine", "DFineForObjectDetection"),
+        ("deim", "DeimForObjectDetection"),
         ("dab-detr", "DabDetrForObjectDetection"),
         ("deformable_detr", "DeformableDetrForObjectDetection"),
         ("deta", "DetaForObjectDetection"),

src/transformers/models/deim/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2025 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 _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_deim import *
+    from .modeling_deim import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/deim/configuration_deim.py

@@ -0,0 +1,425 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/deim/modular_deim.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_deim.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 Baidu Inc and 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.
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ...utils.backbone_utils import verify_backbone_config_arguments
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+
+# TODO: Attribute map assignment logic should be fixed in modular
+# as well as super() call parsing becuase otherwise we cannot re-write args after initialization
+class DeimConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a [`DeimModel`]. It is used to instantiate a deim
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of deim-X-COCO "[sushmanth/DEIM"](https://huggingface.co/sushmanth/DEIM").
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        initializer_range (`float`, *optional*, defaults to 0.01):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_bias_prior_prob (`float`, *optional*):
+            The prior probability used by the bias initializer to initialize biases for `enc_score_head` and `class_embed`.
+            If `None`, `prior_prob` computed as `prior_prob = 1 / (num_labels + 1)` while initializing model weights.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        batch_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the batch normalization layers.
+        backbone_config (`Dict`, *optional*, defaults to `RTDetrResNetConfig()`):
+            The configuration of the backbone model.
+        backbone (`str`, *optional*):
+            Name of backbone to use when `backbone_config` is `None`. If `use_pretrained_backbone` is `True`, this
+            will load the corresponding pretrained weights from the timm or transformers library. If `use_pretrained_backbone`
+            is `False`, this loads the backbone's config and uses that to initialize the backbone with random weights.
+        use_pretrained_backbone (`bool`, *optional*, defaults to `False`):
+            Whether to use pretrained weights for the backbone.
+        use_timm_backbone (`bool`, *optional*, defaults to `False`):
+            Whether to load `backbone` from the timm library. If `False`, the backbone is loaded from the transformers
+            library.
+        freeze_backbone_batch_norms (`bool`, *optional*, defaults to `True`):
+            Whether to freeze the batch normalization layers in the backbone.
+        backbone_kwargs (`dict`, *optional*):
+            Keyword arguments to be passed to AutoBackbone when loading from a checkpoint
+            e.g. `{'out_indices': (0, 1, 2, 3)}`. Cannot be specified if `backbone_config` is set.
+        encoder_hidden_dim (`int`, *optional*, defaults to 256):
+            Dimension of the layers in hybrid encoder.
+        encoder_in_channels (`list`, *optional*, defaults to `[512, 1024, 2048]`):
+            Multi level features input for encoder.
+        feat_strides (`List[int]`, *optional*, defaults to `[8, 16, 32]`):
+            Strides used in each feature map.
+        encoder_layers (`int`, *optional*, defaults to 1):
+            Total of layers to be used by the encoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The ratio for all dropout layers.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        encode_proj_layers (`List[int]`, *optional*, defaults to `[2]`):
+            Indexes of the projected layers to be used in the encoder.
+        positional_encoding_temperature (`int`, *optional*, defaults to 10000):
+            The temperature parameter used to create the positional encodings.
+        encoder_activation_function (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        activation_function (`str`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the general layer. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        eval_size (`Tuple[int, int]`, *optional*):
+            Height and width used to computes the effective height and width of the position embeddings after taking
+            into account the stride.
+        normalize_before (`bool`, *optional*, defaults to `False`):
+            Determine whether to apply layer normalization in the transformer encoder layer before self-attention and
+            feed-forward modules.
+        hidden_expansion (`float`, *optional*, defaults to 1.0):
+            Expansion ratio to enlarge the dimension size of RepVGGBlock and CSPRepLayer.
+        d_model (`int`, *optional*, defaults to 256):
+            Dimension of the layers exclude hybrid encoder.
+        num_queries (`int`, *optional*, defaults to 300):
+            Number of object queries.
+        decoder_in_channels (`list`, *optional*, defaults to `[256, 256, 256]`):
+            Multi level features dimension for decoder
+        decoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        num_feature_levels (`int`, *optional*, defaults to 3):
+            The number of input feature levels.
+        decoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the decoder.
+        decoder_layers (`int`, *optional*, defaults to 6):
+            Number of decoder layers.
+        decoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_activation_function (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the decoder. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        num_denoising (`int`, *optional*, defaults to 100):
+            The total number of denoising tasks or queries to be used for contrastive denoising.
+        label_noise_ratio (`float`, *optional*, defaults to 0.5):
+            The fraction of denoising labels to which random noise should be added.
+        box_noise_scale (`float`, *optional*, defaults to 1.0):
+            Scale or magnitude of noise to be added to the bounding boxes.
+        learn_initial_query (`bool`, *optional*, defaults to `False`):
+            Indicates whether the initial query embeddings for the decoder should be learned during training
+        anchor_image_size (`Tuple[int, int]`, *optional*):
+            Height and width of the input image used during evaluation to generate the bounding box anchors. If None, automatic generate anchor is applied.
+        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.
+        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
+            Whether the architecture has an encoder decoder structure.
+        matcher_alpha (`float`, *optional*, defaults to 0.25):
+            Parameter alpha used by the Hungarian Matcher.
+        matcher_gamma (`float`, *optional*, defaults to 2.0):
+            Parameter gamma used by the Hungarian Matcher.
+        matcher_class_cost (`float`, *optional*, defaults to 2.0):
+            The relative weight of the class loss used by the Hungarian Matcher.
+        matcher_bbox_cost (`float`, *optional*, defaults to 5.0):
+            The relative weight of the bounding box loss used by the Hungarian Matcher.
+        matcher_giou_cost (`float`, *optional*, defaults to 2.0):
+            The relative weight of the giou loss of used by the Hungarian Matcher.
+        use_focal_loss (`bool`, *optional*, defaults to `True`):
+            Parameter informing if focal focal should be used.
+        auxiliary_loss (`bool`, *optional*, defaults to `True`):
+            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
+        focal_loss_alpha (`float`, *optional*, defaults to 0.75):
+            Parameter alpha used to compute the focal loss.
+        focal_loss_gamma (`float`, *optional*, defaults to 2.0):
+            Parameter gamma used to compute the focal loss.
+        weight_loss_vfl (`float`, *optional*, defaults to 1.0):
+            Relative weight of the varifocal loss in the object detection loss.
+        weight_loss_bbox (`float`, *optional*, defaults to 5.0):
+            Relative weight of the L1 bounding box loss in the object detection loss.
+        weight_loss_giou (`float`, *optional*, defaults to 2.0):
+            Relative weight of the generalized IoU loss in the object detection loss.
+        weight_loss_fgl (`float`, *optional*, defaults to 0.15):
+            Relative weight of the fine-grained localization loss in the object detection loss.
+        weight_loss_ddf (`float`, *optional*, defaults to 1.5):
+            Relative weight of the decoupled distillation focal loss in the object detection loss.
+        eos_coefficient (`float`, *optional*, defaults to 0.0001):
+            Relative classification weight of the 'no-object' class in the object detection loss.
+        eval_idx (`int`, *optional*, defaults to -1):
+            Index of the decoder layer to use for evaluation. If negative, counts from the end
+            (e.g., -1 means use the last layer). This allows for early prediction in the decoder
+            stack while still training later layers.
+        layer_scale (`float`, *optional*, defaults to `1.0`):
+            Scaling factor for the hidden dimension in later decoder layers. Used to adjust the
+            model capacity after the evaluation layer.
+        max_num_bins (`int`, *optional*, defaults to 32):
+            Maximum number of bins for the distribution-guided bounding box refinement.
+            Higher values allow for more fine-grained localization but increase computation.
+        reg_scale (`float`, *optional*, defaults to 4.0):
+            Scale factor for the regression distribution. Controls the range and granularity
+            of the bounding box refinement process.
+        depth_mult (`float`, *optional*, defaults to 1.0):
+            Multiplier for the number of blocks in RepNCSPELAN4 layers. Used to scale the model's
+            depth while maintaining its architecture.
+        top_prob_values (`int`, *optional*, defaults to 4):
+            Number of top probability values to consider from each corner's distribution.
+        lqe_hidden_dim (`int`, *optional*, defaults to 64):
+            Hidden dimension size for the Location Quality Estimator (LQE) network.
+        lqe_layers (`int`, *optional*, defaults to 2):
+            Number of layers in the Location Quality Estimator MLP.
+        decoder_offset_scale (`float`, *optional*, defaults to 0.5):
+            Offset scale used in deformable attention.
+        decoder_method (`str`, *optional*, defaults to `"default"`):
+            The method to use for the decoder: `"default"` or `"discrete"`.
+        up (`float`, *optional*, defaults to 0.5):
+            Controls the upper bounds of the Weighting Function.
+    """
+
+    model_type = "deim"
+    layer_types = ["basic", "bottleneck"]
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(
+        self,
+        initializer_range=0.01,
+        initializer_bias_prior_prob=None,
+        layer_norm_eps=1e-5,
+        batch_norm_eps=1e-5,
+        # backbone
+        backbone_config=None,
+        backbone=None,
+        use_pretrained_backbone=False,
+        use_timm_backbone=False,
+        freeze_backbone_batch_norms=True,
+        backbone_kwargs=None,
+        # encoder HybridEncoder
+        encoder_hidden_dim=256,
+        encoder_in_channels=[512, 1024, 2048],
+        feat_strides=[8, 16, 32],
+        encoder_layers=1,
+        encoder_ffn_dim=1024,
+        encoder_attention_heads=8,
+        dropout=0.0,
+        activation_dropout=0.0,
+        encode_proj_layers=[2],
+        positional_encoding_temperature=10000,
+        encoder_activation_function="gelu",
+        activation_function="silu",
+        eval_size=None,
+        normalize_before=False,
+        hidden_expansion=1.0,
+        # decoder DeimTransformer
+        d_model=256,
+        num_queries=300,
+        decoder_in_channels=[256, 256, 256],
+        decoder_ffn_dim=1024,
+        num_feature_levels=3,
+        decoder_n_points=4,
+        decoder_layers=6,
+        decoder_attention_heads=8,
+        decoder_activation_function="relu",
+        attention_dropout=0.0,
+        num_denoising=100,
+        label_noise_ratio=0.5,
+        box_noise_scale=1.0,
+        learn_initial_query=False,
+        anchor_image_size=None,
+        with_box_refine=True,
+        is_encoder_decoder=True,
+        # Loss
+        matcher_alpha=0.25,
+        matcher_gamma=2.0,
+        matcher_class_cost=2.0,
+        matcher_bbox_cost=5.0,
+        matcher_giou_cost=2.0,
+        use_focal_loss=True,
+        auxiliary_loss=True,
+        focal_loss_alpha=0.75,
+        focal_loss_gamma=2.0,
+        weight_loss_vfl=1.0,
+        weight_loss_bbox=5.0,
+        weight_loss_giou=2.0,
+        weight_loss_fgl=0.15,
+        weight_loss_ddf=1.5,
+        eos_coefficient=1e-4,
+        eval_idx=-1,
+        layer_scale=1,
+        max_num_bins=32,
+        reg_scale=4.0,
+        depth_mult=1.0,
+        top_prob_values=4,
+        lqe_hidden_dim=64,
+        lqe_layers=2,
+        decoder_offset_scale=0.5,
+        decoder_method="default",
+        up=0.5,
+        **kwargs,
+    ):
+        self.initializer_range = initializer_range
+        self.initializer_bias_prior_prob = initializer_bias_prior_prob
+        self.layer_norm_eps = layer_norm_eps
+        self.batch_norm_eps = batch_norm_eps
+        # backbone
+        if backbone_config is None and backbone is None:
+            logger.info(
+                "`backbone_config` and `backbone` are `None`. Initializing the config with the default `HGNet-V2` backbone."
+            )
+            backbone_model_type = "hgnet_v2"
+            config_class = CONFIG_MAPPING[backbone_model_type]
+            # this will map it to RTDetrResNetConfig
+            # note: we can instead create HGNetV2Config
+            # and we would need to create HGNetV2Backbone
+            backbone_config = config_class(
+                num_channels=3,
+                embedding_size=64,
+                hidden_sizes=[256, 512, 1024, 2048],
+                depths=[3, 4, 6, 3],
+                layer_type="bottleneck",
+                hidden_act="relu",
+                downsample_in_first_stage=False,
+                downsample_in_bottleneck=False,
+                out_features=None,
+                out_indices=[2, 3, 4],
+            )
+        elif 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)
+
+        verify_backbone_config_arguments(
+            use_timm_backbone=use_timm_backbone,
+            use_pretrained_backbone=use_pretrained_backbone,
+            backbone=backbone,
+            backbone_config=backbone_config,
+            backbone_kwargs=backbone_kwargs,
+        )
+
+        self.backbone_config = backbone_config
+        self.backbone = backbone
+        self.use_pretrained_backbone = use_pretrained_backbone
+        self.use_timm_backbone = use_timm_backbone
+        self.freeze_backbone_batch_norms = freeze_backbone_batch_norms
+        self.backbone_kwargs = backbone_kwargs
+        # encoder
+        self.encoder_hidden_dim = encoder_hidden_dim
+        self.encoder_in_channels = encoder_in_channels
+        self.feat_strides = feat_strides
+        self.encoder_attention_heads = encoder_attention_heads
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+        self.encode_proj_layers = encode_proj_layers
+        self.encoder_layers = encoder_layers
+        self.positional_encoding_temperature = positional_encoding_temperature
+        self.eval_size = eval_size
+        self.normalize_before = normalize_before
+        self.encoder_activation_function = encoder_activation_function
+        self.activation_function = activation_function
+        self.hidden_expansion = hidden_expansion
+        # decoder
+        self.d_model = d_model
+        self.num_queries = num_queries
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_in_channels = decoder_in_channels
+        self.num_feature_levels = num_feature_levels
+        self.decoder_n_points = decoder_n_points
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_activation_function = decoder_activation_function
+        self.attention_dropout = attention_dropout
+        self.num_denoising = num_denoising
+        self.label_noise_ratio = label_noise_ratio
+        self.box_noise_scale = box_noise_scale
+        self.learn_initial_query = learn_initial_query
+        self.anchor_image_size = anchor_image_size
+        self.auxiliary_loss = auxiliary_loss
+        self.with_box_refine = with_box_refine
+        # Loss
+        self.matcher_alpha = matcher_alpha
+        self.matcher_gamma = matcher_gamma
+        self.matcher_class_cost = matcher_class_cost
+        self.matcher_bbox_cost = matcher_bbox_cost
+        self.matcher_giou_cost = matcher_giou_cost
+        self.use_focal_loss = use_focal_loss
+        self.focal_loss_alpha = focal_loss_alpha
+        self.focal_loss_gamma = focal_loss_gamma
+        self.weight_loss_vfl = weight_loss_vfl
+        self.weight_loss_bbox = weight_loss_bbox
+        self.weight_loss_giou = weight_loss_giou
+        self.weight_loss_fgl = weight_loss_fgl
+        self.weight_loss_ddf = weight_loss_ddf
+        self.eos_coefficient = eos_coefficient
+        # add the new attributes with the given values or defaults
+        self.eval_idx = eval_idx
+        self.layer_scale = layer_scale
+        self.max_num_bins = max_num_bins
+        self.reg_scale = reg_scale
+        self.depth_mult = depth_mult
+        self.decoder_offset_scale = decoder_offset_scale
+        self.decoder_method = decoder_method
+        self.top_prob_values = top_prob_values
+        self.lqe_hidden_dim = lqe_hidden_dim
+        self.lqe_layers = lqe_layers
+        self.up = up
+
+        if isinstance(self.decoder_n_points, list):
+            if len(self.decoder_n_points) != self.num_feature_levels:
+                raise ValueError(
+                    f"Length of decoder_n_points list ({len(self.decoder_n_points)}) must match num_feature_levels ({self.num_feature_levels})."
+                )
+
+        head_dim = self.d_model // self.decoder_attention_heads
+        if head_dim * self.decoder_attention_heads != self.d_model:
+            raise ValueError(
+                f"Embedded dimension {self.d_model} must be divisible by decoder_attention_heads {self.decoder_attention_heads}"
+            )
+        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
+
+    @classmethod
+    def from_backbone_configs(cls, backbone_config: PretrainedConfig, **kwargs):
+        """Instantiate a [`DeimConfig`] (or a derived class) from a pre-trained backbone model configuration and DETR model
+        configuration.
+
+            Args:
+                backbone_config ([`PretrainedConfig`]):
+                    The backbone configuration.
+
+            Returns:
+                [`DeimConfig`]: An instance of a configuration object
+        """
+        return cls(
+            backbone_config=backbone_config,
+            **kwargs,
+        )
+
+
+__all__ = ["DeimConfig"]

src/transformers/models/deim/convert_deim_original_pytorch_checkpoint_to_hf.py

@@ -0,0 +1,688 @@
+# coding=utf-8
+# Copyright 2025 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.
+
+import argparse
+import json
+import re
+from pathlib import Path
+
+import requests
+import torch
+from huggingface_hub import hf_hub_download
+from PIL import Image
+from torchvision import transforms
+
+from transformers import DeimConfig, DeimForObjectDetection, RTDetrImageProcessor
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_deim_config(model_name: str) -> DeimConfig:
+    config = DeimConfig()
+
+    config.num_labels = 80
+    repo_id = "huggingface/label-files"
+    filename = "object365-id2label.json" if "obj365" in model_name else "coco-detection-mmdet-id2label.json"
+    id2label = json.load(open(hf_hub_download(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()}
+
+    config.backbone_config.hidden_sizes = [64, 128, 256, 512]
+    config.backbone_config.layer_type = "basic"
+    config.backbone_config.embedding_size = 32
+    config.hidden_expansion = 1.0
+    config.decoder_layers = 6
+
+    if model_name in ["dfine_x_coco", "dfine_x_obj2coco", "dfine_x_obj365"]:
+        config.backbone_config.hidden_sizes = [256, 512, 1024, 2048]
+        config.backbone_config.stage_in_channels = [64, 128, 512, 1024]
+        config.backbone_config.stage_mid_channels = [64, 128, 256, 512]
+        config.backbone_config.stage_out_channels = [128, 512, 1024, 2048]
+        config.backbone_config.stage_num_blocks = [1, 2, 5, 2]
+        config.backbone_config.stage_downsample = [False, True, True, True]
+        config.backbone_config.stage_light_block = [False, False, True, True]
+        config.backbone_config.stage_kernel_size = [3, 3, 5, 5]
+        config.backbone_config.stage_numb_of_layers = [6, 6, 6, 6]
+        config.backbone_config.stem_channels = [3, 32, 64]
+        config.encoder_in_channels = [512, 1024, 2048]
+        config.encoder_hidden_dim = 384
+        config.encoder_ffn_dim = 2048
+        config.decoder_n_points = [3, 6, 3]
+        config.decoder_in_channels = [384, 384, 384]
+        if model_name == "dfine_x_obj365":
+            config.num_labels = 366
+    elif model_name in ["dfine_m_coco", "dfine_m_obj2coco", "dfine_m_obj365"]:
+        config.backbone_config.hidden_sizes = [192, 384, 768, 1536]
+        config.backbone_config.stem_channels = [3, 24, 32]
+        config.backbone_config.stage_in_channels = [32, 96, 384, 768]
+        config.backbone_config.stage_mid_channels = [32, 64, 128, 256]
+        config.backbone_config.stage_out_channels = [96, 384, 768, 1536]
+        config.backbone_config.stage_num_blocks = [1, 1, 3, 1]
+        config.backbone_config.stage_downsample = [False, True, True, True]
+        config.backbone_config.stage_light_block = [False, False, True, True]
+        config.backbone_config.stage_kernel_size = [3, 3, 5, 5]
+        config.backbone_config.stage_numb_of_layers = [4, 4, 4, 4]
+        config.decoder_layers = 4
+        config.decoder_n_points = [3, 6, 3]
+        config.encoder_in_channels = [384, 768, 1536]
+        config.backbone_config.use_learnable_affine_block = True
+        config.depth_mult = 0.67
+        if model_name == "dfine_m_obj365":
+            config.num_labels = 366
+    elif model_name in ["dfine_l_coco", "dfine_l_obj2coco_e25", "dfine_l_obj365"]:
+        config.backbone_config.hidden_sizes = [256, 512, 1024, 2048]
+        config.backbone_config.stem_channels = [3, 32, 48]
+        config.backbone_config.stage_in_channels = [48, 128, 512, 1024]
+        config.backbone_config.stage_mid_channels = [48, 96, 192, 384]
+        config.backbone_config.stage_out_channels = [128, 512, 1024, 2048]
+        config.backbone_config.stage_num_blocks = [1, 1, 3, 1]
+        config.backbone_config.stage_downsample = [False, True, True, True]
+        config.backbone_config.stage_light_block = [False, False, True, True]
+        config.backbone_config.stage_kernel_size = [3, 3, 5, 5]
+        config.backbone_config.stage_numb_of_layers = [6, 6, 6, 6]
+        config.encoder_ffn_dim = 1024
+        config.encoder_in_channels = [512, 1024, 2048]
+        config.decoder_n_points = [3, 6, 3]
+        if model_name == "dfine_l_obj365":
+            config.num_labels = 366
+    elif model_name in ["dfine_n_coco", "dfine_n_obj2coco_e25", "dfine_n_obj365"]:
+        config.backbone_config.hidden_sizes = [128, 256, 512, 1024]
+        config.backbone_config.stem_channels = [3, 16, 16]
+        config.backbone_config.stage_in_channels = [16, 64, 256, 512]
+        config.backbone_config.stage_mid_channels = [16, 32, 64, 128]
+        config.backbone_config.stage_out_channels = [64, 256, 512, 1024]
+        config.backbone_config.stage_num_blocks = [1, 1, 2, 1]
+        config.backbone_config.stage_downsample = [False, True, True, True]
+        config.backbone_config.stage_light_block = [False, False, True, True]
+        config.backbone_config.stage_kernel_size = [3, 3, 5, 5]
+        config.backbone_config.stage_numb_of_layers = [3, 3, 3, 3]
+        config.backbone_config.out_indices = [3, 4]
+        config.backbone_config.use_learnable_affine_block = True
+        config.num_feature_levels = 2
+        config.encoder_ffn_dim = 512
+        config.encode_proj_layers = [1]
+        config.d_model = 128
+        config.encoder_hidden_dim = 128
+        config.decoder_ffn_dim = 512
+        config.encoder_in_channels = [512, 1024]
+        config.decoder_n_points = [6, 6]
+        config.decoder_in_channels = [128, 128]
+        config.feat_strides = [16, 32]
+        config.depth_mult = 0.5
+        config.decoder_layers = 3
+        config.hidden_expansion = 0.34
+        if model_name == "dfine_n_obj365":
+            config.num_labels = 366
+    else:
+        config.backbone_config.hidden_sizes = [128, 256, 512, 1024]
+        config.backbone_config.stem_channels = [3, 16, 16]
+        config.backbone_config.stage_in_channels = [16, 64, 256, 512]
+        config.backbone_config.stage_mid_channels = [16, 32, 64, 128]
+        config.backbone_config.stage_out_channels = [64, 256, 512, 1024]
+        config.backbone_config.stage_num_blocks = [1, 1, 2, 1]
+        config.backbone_config.stage_downsample = [False, True, True, True]
+        config.backbone_config.stage_light_block = [False, False, True, True]
+        config.backbone_config.stage_kernel_size = [3, 3, 5, 5]
+        config.backbone_config.stage_numb_of_layers = [3, 3, 3, 3]
+        config.decoder_layers = 3
+        config.hidden_expansion = 0.5
+        config.depth_mult = 0.34
+        config.decoder_n_points = [3, 6, 3]
+        config.encoder_in_channels = [256, 512, 1024]
+        config.backbone_config.use_learnable_affine_block = True
+        if model_name == "dfine_s_obj365":
+            config.num_labels = 366
+
+    return config
+
+
+def load_original_state_dict(repo_id, model_name):
+    directory_path = hf_hub_download(repo_id=repo_id, filename=f"{model_name}.pth")
+
+    original_state_dict = {}
+    model = torch.load(directory_path, map_location="cpu")["model"]
+    for key in model.keys():
+        original_state_dict[key] = model[key]
+
+    return original_state_dict
+
+
+ORIGINAL_TO_CONVERTED_KEY_MAPPING = {
+    # Decoder base mappings
+    r"decoder.valid_mask": r"model.decoder.valid_mask",
+    r"decoder.anchors": r"model.decoder.anchors",
+    r"decoder.up": r"model.decoder.up",
+    r"decoder.reg_scale": r"model.decoder.reg_scale",
+    # Backbone stem mappings - including stem2a and stem2b
+    r"backbone.stem.stem1.conv.weight": r"model.backbone.model.embedder.stem1.convolution.weight",
+    r"backbone.stem.stem2a.conv.weight": r"model.backbone.model.embedder.stem2a.convolution.weight",
+    r"backbone.stem.stem2b.conv.weight": r"model.backbone.model.embedder.stem2b.convolution.weight",
+    r"backbone.stem.stem3.conv.weight": r"model.backbone.model.embedder.stem3.convolution.weight",
+    r"backbone.stem.stem4.conv.weight": r"model.backbone.model.embedder.stem4.convolution.weight",
+    # Stem normalization
+    r"backbone.stem.stem1.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.embedder.stem1.normalization.\1",
+    r"backbone.stem.stem2a.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.embedder.stem2a.normalization.\1",
+    r"backbone.stem.stem2b.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.embedder.stem2b.normalization.\1",
+    r"backbone.stem.stem3.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.embedder.stem3.normalization.\1",
+    r"backbone.stem.stem4.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.embedder.stem4.normalization.\1",
+    # Stem lab parameters - fixed with .lab in the path
+    r"backbone.stem.stem1.lab.(scale|bias)": r"model.backbone.model.embedder.stem1.lab.\1",
+    r"backbone.stem.stem2a.lab.(scale|bias)": r"model.backbone.model.embedder.stem2a.lab.\1",
+    r"backbone.stem.stem2b.lab.(scale|bias)": r"model.backbone.model.embedder.stem2b.lab.\1",
+    r"backbone.stem.stem3.lab.(scale|bias)": r"model.backbone.model.embedder.stem3.lab.\1",
+    r"backbone.stem.stem4.lab.(scale|bias)": r"model.backbone.model.embedder.stem4.lab.\1",
+    # Backbone stages mappings
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv.weight": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.convolution.weight",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.normalization.\4",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv1.conv.weight": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv1.convolution.weight",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv2.conv.weight": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv2.convolution.weight",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv1.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv1.normalization.\4",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv2.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv2.normalization.\4",
+    # Backbone stages aggregation
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.0.conv.weight": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.0.convolution.weight",
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.1.conv.weight": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.1.convolution.weight",
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.0.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.0.normalization.\3",
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.1.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.1.normalization.\3",
+    # Backbone stages lab parameters for aggregation
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.0.lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.0.lab.\3",
+    r"backbone.stages.(\d+).blocks.(\d+).aggregation.1.lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.blocks.\2.aggregation.1.lab.\3",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.lab.\4",
+    # Conv1/Conv2 layers with lab
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv1.lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv1.lab.\4",
+    r"backbone.stages.(\d+).blocks.(\d+).layers.(\d+).conv2.lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.blocks.\2.layers.\3.conv2.lab.\4",
+    # Downsample with lab
+    r"backbone.stages.(\d+).downsample.lab.(scale|bias)": r"model.backbone.model.encoder.stages.\1.downsample.lab.\2",
+    # Backbone downsample
+    r"backbone.stages.(\d+).downsample.conv.weight": r"model.backbone.model.encoder.stages.\1.downsample.convolution.weight",
+    r"backbone.stages.(\d+).downsample.bn.(weight|bias|running_mean|running_var)": r"model.backbone.model.encoder.stages.\1.downsample.normalization.\2",
+    # Encoder mappings
+    r"encoder.encoder.(\d+).layers.0.self_attn.out_proj.(weight|bias)": r"model.encoder.encoder.\1.layers.0.self_attn.out_proj.\2",
+    r"encoder.encoder.(\d+).layers.0.linear1.(weight|bias)": r"model.encoder.encoder.\1.layers.0.fc1.\2",
+    r"encoder.encoder.(\d+).layers.0.linear2.(weight|bias)": r"model.encoder.encoder.\1.layers.0.fc2.\2",
+    r"encoder.encoder.(\d+).layers.0.norm1.(weight|bias)": r"model.encoder.encoder.\1.layers.0.self_attn_layer_norm.\2",
+    r"encoder.encoder.(\d+).layers.0.norm2.(weight|bias)": r"model.encoder.encoder.\1.layers.0.final_layer_norm.\2",
+    # Encoder projections and convolutions
+    r"encoder.input_proj.(\d+).conv.weight": r"model.encoder_input_proj.\1.0.weight",
+    r"encoder.input_proj.(\d+).norm.(weight|bias|running_mean|running_var)": r"model.encoder_input_proj.\1.1.\2",
+    r"encoder.lateral_convs.(\d+).conv.weight": r"model.encoder.lateral_convs.\1.conv.weight",
+    r"encoder.lateral_convs.(\d+).norm.(weight|bias|running_mean|running_var)": r"model.encoder.lateral_convs.\1.norm.\2",
+    # FPN blocks - complete structure
+    # Basic convolutions
+    r"encoder.fpn_blocks.(\d+).cv1.conv.weight": r"model.encoder.fpn_blocks.\1.conv1.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.conv1.norm.\2",
+    # CSP Rep1 path
+    r"encoder.fpn_blocks.(\d+).cv2.0.conv1.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep1.conv1.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv2.0.conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep1.conv1.norm.\2",
+    r"encoder.fpn_blocks.(\d+).cv2.0.conv2.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep1.conv2.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv2.0.conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep1.conv2.norm.\2",
+    r"encoder.fpn_blocks.(\d+).cv2.1.conv.weight": r"model.encoder.fpn_blocks.\1.conv2.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv2.1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.conv2.norm.\2",
+    # CSP Rep2 path
+    r"encoder.fpn_blocks.(\d+).cv3.0.conv1.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep2.conv1.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv3.0.conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep2.conv1.norm.\2",
+    r"encoder.fpn_blocks.(\d+).cv3.0.conv2.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep2.conv2.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv3.0.conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep2.conv2.norm.\2",
+    r"encoder.fpn_blocks.(\d+).cv3.1.conv.weight": r"model.encoder.fpn_blocks.\1.conv3.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv3.1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.conv3.norm.\2",
+    # Final conv
+    r"encoder.fpn_blocks.(\d+).cv4.conv.weight": r"model.encoder.fpn_blocks.\1.conv4.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv4.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.conv4.norm.\2",
+    # Bottlenecks for CSP Rep1
+    r"encoder.fpn_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv1.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep1.bottlenecks.\2.conv1.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep1.bottlenecks.\2.conv1.norm.\3",
+    r"encoder.fpn_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv2.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep1.bottlenecks.\2.conv2.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep1.bottlenecks.\2.conv2.norm.\3",
+    # Bottlenecks for CSP Rep2
+    r"encoder.fpn_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv1.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep2.bottlenecks.\2.conv1.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep2.bottlenecks.\2.conv1.norm.\3",
+    r"encoder.fpn_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv2.conv.weight": r"model.encoder.fpn_blocks.\1.csp_rep2.bottlenecks.\2.conv2.conv.weight",
+    r"encoder.fpn_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.fpn_blocks.\1.csp_rep2.bottlenecks.\2.conv2.norm.\3",
+    # PAN blocks - complete structure
+    # Basic convolutions
+    r"encoder.pan_blocks.(\d+).cv1.conv.weight": r"model.encoder.pan_blocks.\1.conv1.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.conv1.norm.\2",
+    # CSP Rep1 path
+    r"encoder.pan_blocks.(\d+).cv2.0.conv1.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep1.conv1.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv2.0.conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep1.conv1.norm.\2",
+    r"encoder.pan_blocks.(\d+).cv2.0.conv2.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep1.conv2.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv2.0.conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep1.conv2.norm.\2",
+    r"encoder.pan_blocks.(\d+).cv2.1.conv.weight": r"model.encoder.pan_blocks.\1.conv2.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv2.1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.conv2.norm.\2",
+    # CSP Rep2 path
+    r"encoder.pan_blocks.(\d+).cv3.0.conv1.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep2.conv1.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv3.0.conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep2.conv1.norm.\2",
+    r"encoder.pan_blocks.(\d+).cv3.0.conv2.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep2.conv2.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv3.0.conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep2.conv2.norm.\2",
+    r"encoder.pan_blocks.(\d+).cv3.1.conv.weight": r"model.encoder.pan_blocks.\1.conv3.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv3.1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.conv3.norm.\2",
+    # Final conv
+    r"encoder.pan_blocks.(\d+).cv4.conv.weight": r"model.encoder.pan_blocks.\1.conv4.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv4.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.conv4.norm.\2",
+    # Bottlenecks for CSP Rep1
+    r"encoder.pan_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv1.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep1.bottlenecks.\2.conv1.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep1.bottlenecks.\2.conv1.norm.\3",
+    r"encoder.pan_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv2.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep1.bottlenecks.\2.conv2.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv2.0.bottlenecks.(\d+).conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep1.bottlenecks.\2.conv2.norm.\3",
+    # Bottlenecks for CSP Rep2
+    r"encoder.pan_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv1.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep2.bottlenecks.\2.conv1.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv1.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep2.bottlenecks.\2.conv1.norm.\3",
+    r"encoder.pan_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv2.conv.weight": r"model.encoder.pan_blocks.\1.csp_rep2.bottlenecks.\2.conv2.conv.weight",
+    r"encoder.pan_blocks.(\d+).cv3.0.bottlenecks.(\d+).conv2.norm.(weight|bias|running_mean|running_var)": r"model.encoder.pan_blocks.\1.csp_rep2.bottlenecks.\2.conv2.norm.\3",
+    # Downsample convolutions
+    r"encoder.downsample_convs.(\d+).0.cv(\d+).conv.weight": r"model.encoder.downsample_convs.\1.conv\2.conv.weight",
+    r"encoder.downsample_convs.(\d+).0.cv(\d+).norm.(weight|bias|running_mean|running_var)": r"model.encoder.downsample_convs.\1.conv\2.norm.\3",
+    # Decoder layers
+    r"decoder.decoder.layers.(\d+).self_attn.out_proj.(weight|bias)": r"model.decoder.layers.\1.self_attn.out_proj.\2",
+    r"decoder.decoder.layers.(\d+).cross_attn.sampling_offsets.(weight|bias)": r"model.decoder.layers.\1.encoder_attn.sampling_offsets.\2",
+    r"decoder.decoder.layers.(\d+).cross_attn.attention_weights.(weight|bias)": r"model.decoder.layers.\1.encoder_attn.attention_weights.\2",
+    r"decoder.decoder.layers.(\d+).cross_attn.value_proj.(weight|bias)": r"model.decoder.layers.\1.encoder_attn.value_proj.\2",
+    r"decoder.decoder.layers.(\d+).cross_attn.output_proj.(weight|bias)": r"model.decoder.layers.\1.encoder_attn.output_proj.\2",
+    r"decoder.decoder.layers.(\d+).cross_attn.num_points_scale": r"model.decoder.layers.\1.encoder_attn.num_points_scale",
+    r"decoder.decoder.layers.(\d+).gateway.gate.(weight|bias)": r"model.decoder.layers.\1.gateway.gate.\2",
+    r"decoder.decoder.layers.(\d+).gateway.norm.(weight|bias)": r"model.decoder.layers.\1.gateway.norm.\2",
+    r"decoder.decoder.layers.(\d+).norm1.(weight|bias)": r"model.decoder.layers.\1.self_attn_layer_norm.\2",
+    r"decoder.decoder.layers.(\d+).norm2.(weight|bias)": r"model.decoder.layers.\1.encoder_attn_layer_norm.\2",
+    r"decoder.decoder.layers.(\d+).norm3.(weight|bias)": r"model.decoder.layers.\1.final_layer_norm.\2",
+    r"decoder.decoder.layers.(\d+).linear1.(weight|bias)": r"model.decoder.layers.\1.fc1.\2",
+    r"decoder.decoder.layers.(\d+).linear2.(weight|bias)": r"model.decoder.layers.\1.fc2.\2",
+    # LQE layers
+    r"decoder.decoder.lqe_layers.(\d+).reg_conf.layers.(\d+).(weight|bias)": r"model.decoder.lqe_layers.\1.reg_conf.layers.\2.\3",
+    # Decoder heads and projections
+    r"decoder.dec_score_head.(\d+).(weight|bias)": r"model.decoder.class_embed.\1.\2",
+    r"decoder.dec_bbox_head.(\d+).layers.(\d+).(weight|bias)": r"model.decoder.bbox_embed.\1.layers.\2.\3",
+    r"decoder.pre_bbox_head.layers.(\d+).(weight|bias)": r"model.decoder.pre_bbox_head.layers.\1.\2",
+    r"decoder.input_proj.(\d+).conv.weight": r"model.decoder_input_proj.\1.0.weight",
+    r"decoder.input_proj.(\d+).norm.(weight|bias|running_mean|running_var)": r"model.decoder_input_proj.\1.1.\2",
+    # Other decoder components
+    r"decoder.denoising_class_embed.weight": r"model.denoising_class_embed.weight",
+    r"decoder.query_pos_head.layers.(\d+).(weight|bias)": r"model.decoder.query_pos_head.layers.\1.\2",
+    r"decoder.enc_output.proj.(weight|bias)": r"model.enc_output.0.\1",
+    r"decoder.enc_output.norm.(weight|bias)": r"model.enc_output.1.\1",
+    r"decoder.enc_score_head.(weight|bias)": r"model.enc_score_head.\1",
+    r"decoder.enc_bbox_head.layers.(\d+).(weight|bias)": r"model.enc_bbox_head.layers.\1.\2",
+}
+
+
+def convert_old_keys_to_new_keys(state_dict_keys: dict = None):
+    # Use the mapping to rename keys
+    for original_key, converted_key in ORIGINAL_TO_CONVERTED_KEY_MAPPING.items():
+        for key in list(state_dict_keys.keys()):
+            new_key = re.sub(original_key, converted_key, key)
+            if new_key != key:
+                state_dict_keys[new_key] = state_dict_keys.pop(key)
+
+    return state_dict_keys
+
+
+def read_in_q_k_v(state_dict, config, model_name):
+    prefix = ""
+    encoder_hidden_dim = config.encoder_hidden_dim
+
+    # first: transformer encoder
+    for i in range(config.encoder_layers):
+        # read in weights + bias of input projection layer (in PyTorch's MultiHeadAttention, this is a single matrix + bias)
+        in_proj_weight = state_dict.pop(f"{prefix}encoder.encoder.{i}.layers.0.self_attn.in_proj_weight")
+        in_proj_bias = state_dict.pop(f"{prefix}encoder.encoder.{i}.layers.0.self_attn.in_proj_bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.q_proj.weight"] = in_proj_weight[
+            :encoder_hidden_dim, :
+        ]
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.q_proj.bias"] = in_proj_bias[:encoder_hidden_dim]
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.k_proj.weight"] = in_proj_weight[
+            encoder_hidden_dim : 2 * encoder_hidden_dim, :
+        ]
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.k_proj.bias"] = in_proj_bias[
+            encoder_hidden_dim : 2 * encoder_hidden_dim
+        ]
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.v_proj.weight"] = in_proj_weight[
+            -encoder_hidden_dim:, :
+        ]
+        state_dict[f"model.encoder.encoder.{i}.layers.0.self_attn.v_proj.bias"] = in_proj_bias[-encoder_hidden_dim:]
+    # next: transformer decoder (which is a bit more complex because it also includes cross-attention)
+    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"{prefix}decoder.decoder.layers.{i}.self_attn.in_proj_weight", None)
+        in_proj_bias = state_dict.pop(f"{prefix}decoder.decoder.layers.{i}.self_attn.in_proj_bias", None)
+        # next, add query, keys and values (in that order) to the state dict
+        if model_name in ["dfine_n_coco", "dfine_n_obj2coco_e25", "dfine_n_obj365"]:
+            state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:128, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:128]
+            state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[256:384, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[256:384]
+            state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-128:, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-128:]
+        else:
+            state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:256, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:256]
+            state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[256:512, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[256:512]
+            state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-256:, :]
+            state_dict[f"model.decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-256:]
+
+
+# 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_deim_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub, repo_id):
+    """
+    Copy/paste/tweak model's weights to our deim structure.
+    """
+
+    # load default config
+    config = get_deim_config(model_name)
+    state_dict = load_original_state_dict(repo_id, model_name)
+    state_dict.pop("decoder.valid_mask", None)
+    state_dict.pop("decoder.anchors", None)
+    model = DeimForObjectDetection(config)
+    logger.info(f"Converting model {model_name}...")
+
+    state_dict = convert_old_keys_to_new_keys(state_dict)
+    state_dict.pop("decoder.model.decoder.up", None)
+    state_dict.pop("decoder.model.decoder.reg_scale", None)
+
+    # query, key and value matrices need special treatment
+    read_in_q_k_v(state_dict, config, model_name)
+    # important: we need to prepend a prefix to each of the base model keys as the head models use different attributes for them
+    for key in state_dict.copy().keys():
+        if key.endswith("num_batches_tracked"):
+            del state_dict[key]
+        # for two_stage
+        if "bbox_embed" in key or ("class_embed" in key and "denoising_" not in key):
+            state_dict[key.split("model.decoder.")[-1]] = state_dict[key]
+
+    # finally, create HuggingFace model and load state dict
+    model.load_state_dict(state_dict)
+    model.eval()
+
+    # load image processor
+    image_processor = RTDetrImageProcessor()
+
+    # prepare image
+    img = prepare_img()
+
+    # preprocess image
+    transformations = transforms.Compose(
+        [
+            transforms.Resize([640, 640], interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.ToTensor(),
+        ]
+    )
+    original_pixel_values = transformations(img).unsqueeze(0)  # insert batch dimension
+
+    encoding = image_processor(images=img, return_tensors="pt")
+    pixel_values = encoding["pixel_values"]
+
+    assert torch.allclose(original_pixel_values, pixel_values)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model.to(device)
+    pixel_values = pixel_values.to(device)
+
+    outputs = model(pixel_values)
+
+    if model_name == "dfine_x_coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-4.844723, -4.7293096, -4.5971327],
+                [-4.554266, -4.61723, -4.627926],
+                [-4.3934402, -4.6064143, -4.139952],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.2565248, 0.5477609, 0.47644863],
+                [0.7690029, 0.41423926, 0.46148556],
+                [0.1688096, 0.19923759, 0.21118002],
+            ]
+        )
+    elif model_name == "dfine_x_obj2coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-4.230433, -6.6295037, -4.8339615],
+                [-4.085411, -6.3280816, -4.695468],
+                [-3.8968022, -6.336813, -4.67051],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.25707328, 0.54842496, 0.47624254],
+                [0.76967394, 0.41272867, 0.45970756],
+                [0.16882066, 0.19918433, 0.2112098],
+            ]
+        )
+    elif model_name == "dfine_x_obj365":
+        expected_slice_logits = torch.tensor(
+            [
+                [-6.3844957, -3.7549126, -4.6873264],
+                [-5.8433194, -3.4490552, -3.3228905],
+                [-6.5314736, -3.7856622, -4.895984],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.7703046, 0.41329497, 0.45932162],
+                [0.16898105, 0.19876392, 0.21050783],
+                [0.25134972, 0.5517619, 0.4864124],
+            ]
+        )
+    elif model_name == "dfine_m_coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-4.5187078, -4.71708, -4.117749],
+                [-4.513984, -4.937715, -3.829125],
+                [-4.830042, -6.931682, -3.1740026],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.25851426, 0.5489963, 0.4757598],
+                [0.769683, 0.41411665, 0.45988125],
+                [0.16866133, 0.19921188, 0.21207744],
+            ]
+        )
+    elif model_name == "dfine_m_obj2coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-4.520666, -7.6678333, -5.739887],
+                [-4.5053635, -7.510611, -5.452532],
+                [-4.70348, -5.6098466, -5.0199957],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.2567608, 0.5485795, 0.4767465],
+                [0.77035284, 0.41236404, 0.4580645],
+                [0.5498525, 0.27548885, 0.05886984],
+            ]
+        )
+    elif model_name == "dfine_m_obj365":
+        expected_slice_logits = torch.tensor(
+            [
+                [-5.770525, -3.1610885, -5.2807794],
+                [-5.7809954, -3.768266, -5.1146393],
+                [-6.180705, -3.7357295, -3.1651964],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.2529114, 0.5526663, 0.48270613],
+                [0.7712474, 0.41294736, 0.457174],
+                [0.5497157, 0.27588123, 0.05813372],
+            ]
+        )
+    elif model_name == "dfine_l_coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-4.068779, -5.169955, -4.339212],
+                [-3.9461594, -5.0279613, -4.0161457],
+                [-4.218292, -6.196324, -5.175245],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.2564867, 0.5489948, 0.4748876],
+                [0.7693534, 0.4138953, 0.4598034],
+                [0.16875696, 0.19875404, 0.21196914],
+            ]
+        )
+    elif model_name == "dfine_l_obj365":
+        expected_slice_logits = torch.tensor(
+            [
+                [-5.7953215, -3.4901116, -5.4394145],
+                [-5.7032104, -3.671125, -5.76121],
+                [-6.09466, -3.1512096, -4.285499],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.7693825, 0.41265628, 0.4606362],
+                [0.25306237, 0.55187637, 0.4832178],
+                [0.16892478, 0.19880727, 0.21115331],
+            ]
+        )
+    elif model_name == "dfine_l_obj2coco_e25":
+        expected_slice_logits = torch.tensor(
+            [
+                [-3.6098495, -6.633563, -5.1227236],
+                [-3.682696, -6.9178205, -5.414557],
+                [-4.491674, -6.0823426, -4.5718226],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.7697078, 0.41368833, 0.45879585],
+                [0.2573691, 0.54856044, 0.47715297],
+                [0.16895264, 0.19871138, 0.2115552],
+            ]
+        )
+    elif model_name == "dfine_n_coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-3.7827945, -5.0889463, -4.8341026],
+                [-5.3046904, -6.2801714, -2.9276395],
+                [-4.497901, -5.2670407, -6.2380104],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.73334837, 0.4270624, 0.39424777],
+                [0.1680235, 0.1988639, 0.21031213],
+                [0.25370035, 0.5534435, 0.48496848],
+            ]
+        )
+    elif model_name == "dfine_s_coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-3.8097816, -4.7724586, -5.994499],
+                [-5.2974715, -9.499067, -6.1653666],
+                [-5.3502765, -3.9530406, -6.3630295],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.7677696, 0.41479152, 0.46441072],
+                [0.16912134, 0.19869131, 0.2123824],
+                [0.2581653, 0.54818195, 0.47512347],
+            ]
+        )
+    elif model_name == "dfine_s_obj2coco":
+        expected_slice_logits = torch.tensor(
+            [
+                [-6.0208125, -7.532673, -5.0572147],
+                [-3.3595953, -9.057545, -6.376975],
+                [-4.3203554, -9.546032, -6.075504],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.16901012, 0.19883151, 0.21121952],
+                [0.76784194, 0.41266578, 0.46402973],
+                [00.2563128, 0.54797643, 0.47937632],
+            ]
+        )
+    elif model_name == "dfine_s_obj365":
+        expected_slice_logits = torch.tensor(
+            [
+                [-6.3807316, -4.320986, -6.4775343],
+                [-6.5818424, -3.5009093, -5.75824],
+                [-5.748005, -4.3228016, -4.003726],
+            ]
+        )
+        expected_slice_boxes = torch.tensor(
+            [
+                [0.2532072, 0.5491191, 0.48222217],
+                [0.76586807, 0.41175705, 0.46789962],
+                [0.169111, 0.19844547, 0.21069047],
+            ]
+        )
+    else:
+        raise ValueError(f"Unknown deim_name: {model_name}")
+
+    assert torch.allclose(outputs.logits[0, :3, :3], expected_slice_logits.to(outputs.logits.device), atol=1e-3)
+    assert torch.allclose(outputs.pred_boxes[0, :3, :3], expected_slice_boxes.to(outputs.pred_boxes.device), atol=1e-4)
+
+    if pytorch_dump_folder_path is not None:
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        print(f"Saving model {model_name} to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        print(f"Saving image processor to {pytorch_dump_folder_path}")
+        image_processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        # Upload model, image processor and config to the hub
+        logger.info("Uploading PyTorch model and image processor to the hub...")
+        config.push_to_hub(
+            repo_id=repo_id,
+            commit_message="Add config from convert_deim_original_pytorch_checkpoint_to_hf.py",
+        )
+        model.push_to_hub(
+            repo_id=repo_id,
+            commit_message="Add model from convert_deim_original_pytorch_checkpoint_to_hf.py",
+        )
+        image_processor.push_to_hub(
+            repo_id=repo_id,
+            commit_message="Add image processor from convert_deim_original_pytorch_checkpoint_to_hf.py",
+        )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_name",
+        default="dfine_s_coco",
+        type=str,
+        help="model_name of the checkpoint you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
+    )
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether to push the model to the hub or not.")
+    parser.add_argument(
+        "--repo_id",
+        type=str,
+        help="repo_id where the model will be pushed to.",
+    )
+    args = parser.parse_args()
+    convert_deim_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub, args.repo_id)

tests/models/deim/test_modeling_deim.py

@@ -0,0 +1,805 @@
+# coding = utf-8
+# Copyright 2025 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 deim model."""
+
+import inspect
+import math
+import tempfile
+import unittest
+
+from parameterized import parameterized
+
+from transformers import (
+    DeimConfig,
+    HGNetV2Config,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import require_torch, require_torch_gpu, require_vision, slow, torch_device
+from transformers.utils import cached_property
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import DeimForObjectDetection, DeimModel
+
+if is_vision_available():
+    from PIL import Image
+
+from transformers import RTDetrImageProcessor
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+CHECKPOINT = "ustc-community/dfine_s_coco"
+
+
+class DeimModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        is_training=True,
+        use_labels=True,
+        n_targets=3,
+        num_labels=10,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        batch_norm_eps=1e-5,
+        # backbone
+        backbone_config=None,
+        # encoder HybridEncoder
+        encoder_hidden_dim=32,
+        encoder_in_channels=[128, 256, 512],
+        feat_strides=[8, 16, 32],
+        encoder_layers=1,
+        encoder_ffn_dim=64,
+        encoder_attention_heads=2,
+        dropout=0.0,
+        activation_dropout=0.0,
+        encode_proj_layers=[2],
+        positional_encoding_temperature=10000,
+        encoder_activation_function="gelu",
+        activation_function="silu",
+        eval_size=None,
+        normalize_before=False,
+        # decoder DeimTransformer
+        d_model=32,
+        num_queries=30,
+        decoder_in_channels=[32, 32, 32],
+        decoder_ffn_dim=64,
+        num_feature_levels=3,
+        decoder_n_points=[3, 6, 3],
+        decoder_n_levels=3,
+        decoder_layers=2,
+        decoder_attention_heads=2,
+        decoder_activation_function="relu",
+        attention_dropout=0.0,
+        num_denoising=0,
+        label_noise_ratio=0.5,
+        box_noise_scale=1.0,
+        learn_initial_query=False,
+        anchor_image_size=None,
+        image_size=64,
+        disable_custom_kernels=True,
+        with_box_refine=True,
+        decoder_offset_scale=0.5,
+        eval_idx=-1,
+        layer_scale=1,
+        reg_max=32,
+        reg_scale=4.0,
+        depth_mult=0.34,
+        hidden_expansion=0.5,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = 3
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.n_targets = n_targets
+        self.num_labels = num_labels
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.batch_norm_eps = batch_norm_eps
+        self.backbone_config = backbone_config
+        self.encoder_hidden_dim = encoder_hidden_dim
+        self.encoder_in_channels = encoder_in_channels
+        self.feat_strides = feat_strides
+        self.encoder_layers = encoder_layers
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+        self.encode_proj_layers = encode_proj_layers
+        self.positional_encoding_temperature = positional_encoding_temperature
+        self.encoder_activation_function = encoder_activation_function
+        self.activation_function = activation_function
+        self.eval_size = eval_size
+        self.normalize_before = normalize_before
+        self.d_model = d_model
+        self.num_queries = num_queries
+        self.decoder_in_channels = decoder_in_channels
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.num_feature_levels = num_feature_levels
+        self.decoder_n_points = decoder_n_points
+        self.decoder_n_levels = decoder_n_levels
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_activation_function = decoder_activation_function
+        self.attention_dropout = attention_dropout
+        self.decoder_offset_scale = decoder_offset_scale
+        self.eval_idx = eval_idx
+        self.layer_scale = layer_scale
+        self.reg_max = reg_max
+        self.reg_scale = reg_scale
+        self.depth_mult = depth_mult
+        self.num_denoising = num_denoising
+        self.label_noise_ratio = label_noise_ratio
+        self.box_noise_scale = box_noise_scale
+        self.learn_initial_query = learn_initial_query
+        self.anchor_image_size = anchor_image_size
+        self.image_size = image_size
+        self.disable_custom_kernels = disable_custom_kernels
+        self.with_box_refine = with_box_refine
+        self.hidden_expansion = hidden_expansion
+
+        self.encoder_seq_length = math.ceil(self.image_size / 32) * math.ceil(self.image_size / 32)
+
+    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)
+                labels.append(target)
+
+        config = self.get_config()
+        config.num_labels = self.num_labels
+        return config, pixel_values, pixel_mask, labels
+
+    def get_config(self):
+        hidden_sizes = [64, 128, 256, 512]
+        backbone_config = HGNetV2Config(
+            stage_in_channels=[16, 64, 128, 256],
+            stage_mid_channels=[16, 32, 64, 128],
+            stage_out_channels=[64, 128, 256, 512],
+            stage_num_blocks=[1, 1, 2, 1],
+            stage_downsample=[False, True, True, True],
+            stage_light_block=[False, False, True, True],
+            stage_kernel_size=[3, 3, 5, 5],
+            stage_numb_of_layers=[3, 3, 3, 3],
+            embeddings_size=10,
+            hidden_sizes=hidden_sizes,
+            depths=[1, 1, 2, 1],
+            out_features=["stage2", "stage3", "stage4"],
+            out_indices=[2, 3, 4],
+            stem_channels=[3, 16, 16],
+            use_lab=True,
+        )
+        return DeimConfig.from_backbone_configs(
+            backbone_config=backbone_config,
+            encoder_hidden_dim=self.encoder_hidden_dim,
+            encoder_in_channels=self.encoder_in_channels,
+            feat_strides=self.feat_strides,
+            encoder_layers=self.encoder_layers,
+            encoder_ffn_dim=self.encoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            dropout=self.dropout,
+            activation_dropout=self.activation_dropout,
+            encode_proj_layers=self.encode_proj_layers,
+            positional_encoding_temperature=self.positional_encoding_temperature,
+            encoder_activation_function=self.encoder_activation_function,
+            activation_function=self.activation_function,
+            eval_size=self.eval_size,
+            normalize_before=self.normalize_before,
+            d_model=self.d_model,
+            num_queries=self.num_queries,
+            decoder_in_channels=self.decoder_in_channels,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            num_feature_levels=self.num_feature_levels,
+            decoder_n_points=self.decoder_n_points,
+            decoder_n_levels=self.decoder_n_levels,
+            decoder_layers=self.decoder_layers,
+            decoder_attention_heads=self.decoder_attention_heads,
+            decoder_activation_function=self.decoder_activation_function,
+            decoder_offset_scale=self.decoder_offset_scale,
+            eval_idx=self.eval_idx,
+            layer_scale=self.layer_scale,
+            reg_max=self.reg_max,
+            reg_scale=self.reg_scale,
+            depth_mult=self.depth_mult,
+            attention_dropout=self.attention_dropout,
+            num_denoising=self.num_denoising,
+            label_noise_ratio=self.label_noise_ratio,
+            box_noise_scale=self.box_noise_scale,
+            learn_initial_query=self.learn_initial_query,
+            anchor_image_size=self.anchor_image_size,
+            image_size=self.image_size,
+            disable_custom_kernels=self.disable_custom_kernels,
+            with_box_refine=self.with_box_refine,
+        )
+
+    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}
+        return config, inputs_dict
+
+    def create_and_check_deim_model(self, config, pixel_values, pixel_mask, labels):
+        model = DeimModel(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.d_model))
+
+    def create_and_check_deim_object_detection_head_model(self, config, pixel_values, pixel_mask, labels):
+        model = DeimForObjectDetection(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_torch
+class DeimModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (DeimModel, DeimForObjectDetection) if is_torch_available() else ()
+    is_encoder_decoder = True
+    test_torchscript = False
+    test_pruning = False
+    test_head_masking = False
+    test_missing_keys = False
+    test_torch_exportable = True
+
+    # 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__ == "DeimForObjectDetection":
+                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
+                    )
+                    labels.append(target)
+                inputs_dict["labels"] = labels
+
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = DeimModelTester(self)
+        self.config_tester = ConfigTester(
+            self,
+            config_class=DeimConfig,
+            has_text_modality=False,
+            common_properties=["hidden_size", "num_attention_heads"],
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_deim_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deim_model(*config_and_inputs)
+
+    def test_deim_object_detection_head_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deim_object_detection_head_model(*config_and_inputs)
+
+    @unittest.skip(reason="Deim doesn't work well with `nn.DataParallel")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    @unittest.skip(reason="Deim does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Deim does not use test_inputs_embeds_matches_input_ids")
+    def test_inputs_embeds_matches_input_ids(self):
+        pass
+
+    @unittest.skip(reason="Deim does not support input and output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="Deim does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="Deim does not use token embeddings")
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    @unittest.skip(reason="Not relevant for the model")
+    def test_can_init_all_missing_weights(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.encoder_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.encoder_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [
+                    self.model_tester.encoder_attention_heads,
+                    self.model_tester.encoder_seq_length,
+                    self.model_tester.encoder_seq_length,
+                ],
+            )
+            out_len = len(outputs)
+
+            correct_outlen = 15
+
+            # 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__ == "DeimForObjectDetection":
+                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.decoder_layers)
+            self.assertListEqual(
+                list(decoder_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.decoder_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.decoder_layers)
+            self.assertListEqual(
+                list(cross_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_queries,
+                    self.model_tester.decoder_attention_heads,
+                    self.model_tester.decoder_n_levels * self.model_tester.decoder_n_points
+                    if isinstance(self.model_tester.decoder_n_points, int)
+                    else sum(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
+            else:
+                # Deim should maintin encoder_hidden_states output
+                added_hidden_states = 2
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions
+
+            self.assertEqual(len(self_attentions), self.model_tester.encoder_layers)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.encoder_attention_heads,
+                    self.model_tester.encoder_seq_length,
+                    self.model_tester.encoder_seq_length,
+                ],
+            )
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", len(self.model_tester.encoder_in_channels) - 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            self.assertListEqual(
+                list(hidden_states[1].shape[-2:]),
+                [
+                    self.model_tester.image_size // self.model_tester.feat_strides[-1],
+                    self.model_tester.image_size // self.model_tester.feat_strides[-1],
+                ],
+            )
+
+            if config.is_encoder_decoder:
+                hidden_states = outputs.decoder_hidden_states
+
+                expected_num_layers = getattr(
+                    self.model_tester, "expected_num_hidden_layers", self.model_tester.decoder_layers + 1
+                )
+
+                self.assertIsInstance(hidden_states, (list, tuple))
+                self.assertEqual(len(hidden_states), expected_num_layers)
+
+                self.assertListEqual(
+                    list(hidden_states[0].shape[-2:]),
+                    [self.model_tester.num_queries, self.model_tester.d_model],
+                )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    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
+
+        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 first output since last_hidden_state is the first item
+        output = outputs[0]
+
+        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)
+            arg_names = [*signature.parameters.keys()]
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_different_timm_backbone(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # let's pick a random timm backbone
+        config.encoder_in_channels = [24, 40, 432]
+        config.backbone = "tf_mobilenetv3_small_075"
+        config.backbone_config = None
+        config.use_timm_backbone = True
+        config.backbone_kwargs = {"out_indices": [2, 3, 4]}
+
+        for model_class in self.all_model_classes:
+            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 model_class.__name__ == "DeimForObjectDetection":
+                expected_shape = (
+                    self.model_tester.batch_size,
+                    self.model_tester.num_queries,
+                    self.model_tester.num_labels,
+                )
+                self.assertEqual(outputs.logits.shape, expected_shape)
+                # Confirm out_indices was propogated to backbone
+                self.assertEqual(len(model.model.backbone.intermediate_channel_sizes), 3)
+            else:
+                # Confirm out_indices was propogated to backbone
+                self.assertEqual(len(model.backbone.intermediate_channel_sizes), 3)
+
+            self.assertTrue(outputs)
+
+    def test_hf_backbone(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Load a pretrained HF checkpoint as backbone
+        config.backbone = "microsoft/resnet-18"
+        config.backbone_config = None
+        config.use_timm_backbone = False
+        config.use_pretrained_backbone = True
+        config.backbone_kwargs = {"out_indices": [2, 3, 4]}
+
+        for model_class in self.all_model_classes:
+            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 model_class.__name__ == "DeimForObjectDetection":
+                expected_shape = (
+                    self.model_tester.batch_size,
+                    self.model_tester.num_queries,
+                    self.model_tester.num_labels,
+                )
+                self.assertEqual(outputs.logits.shape, expected_shape)
+                # Confirm out_indices was propogated to backbone
+                self.assertEqual(len(model.model.backbone.intermediate_channel_sizes), 3)
+            else:
+                # Confirm out_indices was propogated to backbone
+                self.assertEqual(len(model.backbone.intermediate_channel_sizes), 3)
+
+            self.assertTrue(outputs)
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        configs_no_init.initializer_bias_prior_prob = 0.2
+        bias_value = -1.3863  # log_e ((1 - 0.2) / 0.2)
+
+        failed_cases = []
+
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            # Skip the check for the backbone
+            for name, module in model.named_modules():
+                if module.__class__.__name__ == "DeimConvEncoder":
+                    backbone_params = [f"{name}.{key}" for key in module.state_dict().keys()]
+                    break
+
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if ("class_embed" in name and "bias" in name) or "enc_score_head.bias" in name:
+                        bias_tensor = torch.full_like(param.data, bias_value)
+                        try:
+                            torch.testing.assert_close(param.data, bias_tensor, atol=1e-4, rtol=1e-4)
+                        except AssertionError:
+                            failed_cases.append(
+                                f"Parameter {name} of model {model_class} seems not properly initialized. "
+                                f"Biases should be initialized to {bias_value}, got {param.data}"
+                            )
+                    elif (
+                        "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
+                        or "enc_score_head.weight" in name
+                        or ("class_embed" in name and "weight" in name)
+                        or name in backbone_params
+                    ):
+                        continue
+                    else:
+                        mean = param.data.mean()
+                        round_mean = (mean * 1e9).round() / 1e9
+                        round_mean = round_mean.item()
+                        if round_mean not in [0.0, 1.0]:
+                            failed_cases.append(
+                                f"Parameter {name} of model {model_class} seems not properly initialized. "
+                                f"Mean is {round_mean}, but should be in [0, 1]"
+                            )
+
+        message = "\n" + "\n".join(failed_cases)
+        self.assertTrue(not failed_cases, message)
+
+    @parameterized.expand(["float32", "float16", "bfloat16"])
+    @require_torch_gpu
+    @slow
+    def test_inference_with_different_dtypes(self, torch_dtype_str):
+        torch_dtype = {
+            "float32": torch.float32,
+            "float16": torch.float16,
+            "bfloat16": torch.bfloat16,
+        }[torch_dtype_str]
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device).to(torch_dtype)
+            model.eval()
+            for key, tensor in inputs_dict.items():
+                if tensor.dtype == torch.float32:
+                    inputs_dict[key] = tensor.to(torch_dtype)
+            with torch.no_grad():
+                _ = model(**self._prepare_for_class(inputs_dict, model_class))
+
+    @parameterized.expand(["float32", "float16", "bfloat16"])
+    @require_torch_gpu
+    @slow
+    def test_inference_equivalence_for_static_and_dynamic_anchors(self, torch_dtype_str):
+        torch_dtype = {
+            "float32": torch.float32,
+            "float16": torch.float16,
+            "bfloat16": torch.bfloat16,
+        }[torch_dtype_str]
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        h, w = inputs_dict["pixel_values"].shape[-2:]
+
+        # convert inputs to the desired dtype
+        for key, tensor in inputs_dict.items():
+            if tensor.dtype == torch.float32:
+                inputs_dict[key] = tensor.to(torch_dtype)
+
+        for model_class in self.all_model_classes:
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model_class(config).save_pretrained(tmpdirname)
+                model_static = model_class.from_pretrained(
+                    tmpdirname, anchor_image_size=[h, w], device_map=torch_device, torch_dtype=torch_dtype
+                ).eval()
+                model_dynamic = model_class.from_pretrained(
+                    tmpdirname, anchor_image_size=None, device_map=torch_device, torch_dtype=torch_dtype
+                ).eval()
+
+            self.assertIsNotNone(model_static.config.anchor_image_size)
+            self.assertIsNone(model_dynamic.config.anchor_image_size)
+
+            with torch.no_grad():
+                outputs_static = model_static(**self._prepare_for_class(inputs_dict, model_class))
+                outputs_dynamic = model_dynamic(**self._prepare_for_class(inputs_dict, model_class))
+
+            torch.testing.assert_close(
+                outputs_static.last_hidden_state, outputs_dynamic.last_hidden_state, rtol=1e-4, atol=1e-4
+            )
+
+
+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_torch
+@require_vision
+class DeimModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return RTDetrImageProcessor.from_pretrained(CHECKPOINT) if is_vision_available() else None
+
+    def test_inference_object_detection_head(self):
+        model = DeimForObjectDetection.from_pretrained(CHECKPOINT).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(
+            [
+                [-3.8097816, -4.7724586, -5.994499],
+                [-5.2974715, -9.499067, -6.1653666],
+                [-5.3502765, -3.9530406, -6.3630295],
+            ]
+        ).to(torch_device)
+        expected_boxes = torch.tensor(
+            [
+                [0.7677696, 0.41479152, 0.46441072],
+                [0.16912134, 0.19869131, 0.2123824],
+                [0.2581653, 0.54818195, 0.47512347],
+            ]
+        ).to(torch_device)
+
+        torch.testing.assert_close(outputs.logits[0, :3, :3], expected_logits, atol=1e-4, rtol=1e-4)
+
+        expected_shape_boxes = torch.Size((1, 300, 4))
+        self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
+        torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4, rtol=1e-4)
+
+        # verify postprocessing
+        results = image_processor.post_process_object_detection(
+            outputs, threshold=0.0, target_sizes=[image.size[::-1]]
+        )[0]
+        expected_scores = torch.tensor([0.9642, 0.9542, 0.9536, 0.8548], device=torch_device)
+        expected_labels = [15, 65, 15, 57]
+        expected_slice_boxes = torch.tensor(
+            [
+                [1.3186283e01, 5.4130211e01, 3.1726535e02, 4.7212445e02],
+                [4.0275269e01, 7.2975174e01, 1.7620003e02, 1.1776848e02],
+                [3.4276117e02, 2.3427944e01, 6.3998401e02, 3.7477191e02],
+                [5.8418274e-01, 1.1794567e00, 6.3933154e02, 4.7485995e02],
+            ],
+            device=torch_device,
+        )
+        torch.testing.assert_close(results["scores"][:4], expected_scores, atol=1e-3, rtol=1e-4)
+        self.assertSequenceEqual(results["labels"][:4].tolist(), expected_labels)
+        torch.testing.assert_close(results["boxes"][:4], expected_slice_boxes[:4], atol=1e-3, rtol=1e-4)