Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2024 The HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""Grounding DINO model configuration"""16 17from ...configuration_utils import PretrainedConfig18from ...utils import logging19from ...utils.backbone_utils import verify_backbone_config_arguments20from ..auto import CONFIG_MAPPING21 22 23logger = logging.get_logger(__name__)24 25 26class GroundingDinoConfig(PretrainedConfig):27    r"""28    This is the configuration class to store the configuration of a [`GroundingDinoModel`]. It is used to instantiate a29    Grounding DINO model according to the specified arguments, defining the model architecture. Instantiating a30    configuration with the defaults will yield a similar configuration to that of the Grounding DINO31    [IDEA-Research/grounding-dino-tiny](https://huggingface.co/IDEA-Research/grounding-dino-tiny) architecture.32 33    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the34    documentation from [`PretrainedConfig`] for more information.35 36    Args:37        backbone_config (`PretrainedConfig` or `dict`, *optional*, defaults to `ResNetConfig()`):38            The configuration of the backbone model.39        backbone (`str`, *optional*):40            Name of backbone to use when `backbone_config` is `None`. If `use_pretrained_backbone` is `True`, this41            will load the corresponding pretrained weights from the timm or transformers library. If `use_pretrained_backbone`42            is `False`, this loads the backbone's config and uses that to initialize the backbone with random weights.43        use_pretrained_backbone (`bool`, *optional*, defaults to `False`):44            Whether to use pretrained weights for the backbone.45        use_timm_backbone (`bool`, *optional*, defaults to `False`):46            Whether to load `backbone` from the timm library. If `False`, the backbone is loaded from the transformers47            library.48        backbone_kwargs (`dict`, *optional*):49            Keyword arguments to be passed to AutoBackbone when loading from a checkpoint50            e.g. `{'out_indices': (0, 1, 2, 3)}`. Cannot be specified if `backbone_config` is set.51        text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `BertConfig`):52            The config object or dictionary of the text backbone.53        num_queries (`int`, *optional*, defaults to 900):54            Number of object queries, i.e. detection slots. This is the maximal number of objects55            [`GroundingDinoModel`] can detect in a single image.56        encoder_layers (`int`, *optional*, defaults to 6):57            Number of encoder layers.58        encoder_ffn_dim (`int`, *optional*, defaults to 2048):59            Dimension of the "intermediate" (often named feed-forward) layer in decoder.60        encoder_attention_heads (`int`, *optional*, defaults to 8):61            Number of attention heads for each attention layer in the Transformer encoder.62        decoder_layers (`int`, *optional*, defaults to 6):63            Number of decoder layers.64        decoder_ffn_dim (`int`, *optional*, defaults to 2048):65            Dimension of the "intermediate" (often named feed-forward) layer in decoder.66        decoder_attention_heads (`int`, *optional*, defaults to 8):67            Number of attention heads for each attention layer in the Transformer decoder.68        is_encoder_decoder (`bool`, *optional*, defaults to `True`):69            Whether the model is used as an encoder/decoder or not.70        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):71            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,72            `"relu"`, `"silu"` and `"gelu_new"` are supported.73        d_model (`int`, *optional*, defaults to 256):74            Dimension of the layers.75        dropout (`float`, *optional*, defaults to 0.1):76            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.77        attention_dropout (`float`, *optional*, defaults to 0.0):78            The dropout ratio for the attention probabilities.79        activation_dropout (`float`, *optional*, defaults to 0.0):80            The dropout ratio for activations inside the fully connected layer.81        auxiliary_loss (`bool`, *optional*, defaults to `False`):82            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.83        position_embedding_type (`str`, *optional*, defaults to `"sine"`):84            Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.85        num_feature_levels (`int`, *optional*, defaults to 4):86            The number of input feature levels.87        encoder_n_points (`int`, *optional*, defaults to 4):88            The number of sampled keys in each feature level for each attention head in the encoder.89        decoder_n_points (`int`, *optional*, defaults to 4):90            The number of sampled keys in each feature level for each attention head in the decoder.91        two_stage (`bool`, *optional*, defaults to `True`):92            Whether to apply a two-stage deformable DETR, where the region proposals are also generated by a variant of93            Grounding DINO, which are further fed into the decoder for iterative bounding box refinement.94        class_cost (`float`, *optional*, defaults to 1.0):95            Relative weight of the classification error in the Hungarian matching cost.96        bbox_cost (`float`, *optional*, defaults to 5.0):97            Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.98        giou_cost (`float`, *optional*, defaults to 2.0):99            Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.100        bbox_loss_coefficient (`float`, *optional*, defaults to 5.0):101            Relative weight of the L1 bounding box loss in the object detection loss.102        giou_loss_coefficient (`float`, *optional*, defaults to 2.0):103            Relative weight of the generalized IoU loss in the object detection loss.104        focal_alpha (`float`, *optional*, defaults to 0.25):105            Alpha parameter in the focal loss.106        disable_custom_kernels (`bool`, *optional*, defaults to `False`):107            Disable the use of custom CUDA and CPU kernels. This option is necessary for the ONNX export, as custom108            kernels are not supported by PyTorch ONNX export.109        max_text_len (`int`, *optional*, defaults to 256):110            The maximum length of the text input.111        text_enhancer_dropout (`float`, *optional*, defaults to 0.0):112            The dropout ratio for the text enhancer.113        fusion_droppath (`float`, *optional*, defaults to 0.1):114            The droppath ratio for the fusion module.115        fusion_dropout (`float`, *optional*, defaults to 0.0):116            The dropout ratio for the fusion module.117        embedding_init_target (`bool`, *optional*, defaults to `True`):118            Whether to initialize the target with Embedding weights.119        query_dim (`int`, *optional*, defaults to 4):120            The dimension of the query vector.121        decoder_bbox_embed_share (`bool`, *optional*, defaults to `True`):122            Whether to share the bbox regression head for all decoder layers.123        two_stage_bbox_embed_share (`bool`, *optional*, defaults to `False`):124            Whether to share the bbox embedding between the two-stage bbox generator and the region proposal125            generation.126        positional_embedding_temperature (`float`, *optional*, defaults to 20):127            The temperature for Sine Positional Embedding that is used together with vision backbone.128        init_std (`float`, *optional*, defaults to 0.02):129            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.130        layer_norm_eps (`float`, *optional*, defaults to 1e-05):131            The epsilon used by the layer normalization layers.132 133    Examples:134 135    ```python136    >>> from transformers import GroundingDinoConfig, GroundingDinoModel137 138    >>> # Initializing a Grounding DINO IDEA-Research/grounding-dino-tiny style configuration139    >>> configuration = GroundingDinoConfig()140 141    >>> # Initializing a model (with random weights) from the IDEA-Research/grounding-dino-tiny style configuration142    >>> model = GroundingDinoModel(configuration)143 144    >>> # Accessing the model configuration145    >>> configuration = model.config146    ```"""147 148    model_type = "grounding-dino"149    attribute_map = {150        "hidden_size": "d_model",151        "num_attention_heads": "encoder_attention_heads",152    }153 154    def __init__(155        self,156        backbone_config=None,157        backbone=None,158        use_pretrained_backbone=False,159        use_timm_backbone=False,160        backbone_kwargs=None,161        text_config=None,162        num_queries=900,163        encoder_layers=6,164        encoder_ffn_dim=2048,165        encoder_attention_heads=8,166        decoder_layers=6,167        decoder_ffn_dim=2048,168        decoder_attention_heads=8,169        is_encoder_decoder=True,170        activation_function="relu",171        d_model=256,172        dropout=0.1,173        attention_dropout=0.0,174        activation_dropout=0.0,175        auxiliary_loss=False,176        position_embedding_type="sine",177        num_feature_levels=4,178        encoder_n_points=4,179        decoder_n_points=4,180        two_stage=True,181        class_cost=1.0,182        bbox_cost=5.0,183        giou_cost=2.0,184        bbox_loss_coefficient=5.0,185        giou_loss_coefficient=2.0,186        focal_alpha=0.25,187        disable_custom_kernels=False,188        # other parameters189        max_text_len=256,190        text_enhancer_dropout=0.0,191        fusion_droppath=0.1,192        fusion_dropout=0.0,193        embedding_init_target=True,194        query_dim=4,195        decoder_bbox_embed_share=True,196        two_stage_bbox_embed_share=False,197        positional_embedding_temperature=20,198        init_std=0.02,199        layer_norm_eps=1e-5,200        **kwargs,201    ):202        if backbone_config is None and backbone is None:203            logger.info("`backbone_config` is `None`. Initializing the config with the default `Swin` backbone.")204            backbone_config = CONFIG_MAPPING["swin"](205                window_size=7,206                image_size=224,207                embed_dim=96,208                depths=[2, 2, 6, 2],209                num_heads=[3, 6, 12, 24],210                out_indices=[2, 3, 4],211            )212        elif isinstance(backbone_config, dict):213            backbone_model_type = backbone_config.pop("model_type")214            config_class = CONFIG_MAPPING[backbone_model_type]215            backbone_config = config_class.from_dict(backbone_config)216 217        verify_backbone_config_arguments(218            use_timm_backbone=use_timm_backbone,219            use_pretrained_backbone=use_pretrained_backbone,220            backbone=backbone,221            backbone_config=backbone_config,222            backbone_kwargs=backbone_kwargs,223        )224 225        if text_config is None:226            text_config = {}227            logger.info("text_config is None. Initializing the text config with default values (`BertConfig`).")228 229        self.backbone_config = backbone_config230        self.backbone = backbone231        self.use_pretrained_backbone = use_pretrained_backbone232        self.use_timm_backbone = use_timm_backbone233        self.backbone_kwargs = backbone_kwargs234        self.num_queries = num_queries235        self.d_model = d_model236        self.encoder_ffn_dim = encoder_ffn_dim237        self.encoder_layers = encoder_layers238        self.encoder_attention_heads = encoder_attention_heads239        self.decoder_ffn_dim = decoder_ffn_dim240        self.decoder_layers = decoder_layers241        self.decoder_attention_heads = decoder_attention_heads242        self.dropout = dropout243        self.attention_dropout = attention_dropout244        self.activation_dropout = activation_dropout245        self.activation_function = activation_function246        self.auxiliary_loss = auxiliary_loss247        self.position_embedding_type = position_embedding_type248        # deformable attributes249        self.num_feature_levels = num_feature_levels250        self.encoder_n_points = encoder_n_points251        self.decoder_n_points = decoder_n_points252        self.two_stage = two_stage253        # Hungarian matcher254        self.class_cost = class_cost255        self.bbox_cost = bbox_cost256        self.giou_cost = giou_cost257        # Loss coefficients258        self.bbox_loss_coefficient = bbox_loss_coefficient259        self.giou_loss_coefficient = giou_loss_coefficient260        self.focal_alpha = focal_alpha261        self.disable_custom_kernels = disable_custom_kernels262        # Text backbone263        if isinstance(text_config, dict):264            text_config["model_type"] = text_config.get("model_type", "bert")265            text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)266        elif text_config is None:267            text_config = CONFIG_MAPPING["bert"]()268 269        self.text_config = text_config270        self.max_text_len = max_text_len271 272        # Text Enhancer273        self.text_enhancer_dropout = text_enhancer_dropout274        # Fusion275        self.fusion_droppath = fusion_droppath276        self.fusion_dropout = fusion_dropout277        # Others278        self.embedding_init_target = embedding_init_target279        self.query_dim = query_dim280        self.decoder_bbox_embed_share = decoder_bbox_embed_share281        self.two_stage_bbox_embed_share = two_stage_bbox_embed_share282        if two_stage_bbox_embed_share and not decoder_bbox_embed_share:283            raise ValueError("If two_stage_bbox_embed_share is True, decoder_bbox_embed_share must be True.")284        self.positional_embedding_temperature = positional_embedding_temperature285        self.init_std = init_std286        self.layer_norm_eps = layer_norm_eps287        super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)288 289    @property290    def num_attention_heads(self) -> int:291        return self.encoder_attention_heads292 293    @property294    def hidden_size(self) -> int:295        return self.d_model296 297    @property298    def sub_configs(self):299        sub_configs = {}300        backbone_config = getattr(self, "backbone_config", None)301        text_config = getattr(self, "text_config", None)302        if isinstance(backbone_config, PretrainedConfig):303            sub_configs["backbone_config"] = type(backbone_config)304        if isinstance(text_config, PretrainedConfig):305            sub_configs["text_config"] = type(self.text_config)306        return sub_configs307 308 309__all__ = ["GroundingDinoConfig"]310