Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2022 SHI Labs and 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"""OneFormer model configuration"""16 17from typing import Optional18 19from ...configuration_utils import PretrainedConfig20from ...utils import logging21from ...utils.backbone_utils import verify_backbone_config_arguments22from ..auto import CONFIG_MAPPING23 24 25logger = logging.get_logger(__name__)26 27 28class OneFormerConfig(PretrainedConfig):29    r"""30    This is the configuration class to store the configuration of a [`OneFormerModel`]. It is used to instantiate a31    OneFormer model according to the specified arguments, defining the model architecture. Instantiating a32    configuration with the defaults will yield a similar configuration to that of the OneFormer33    [shi-labs/oneformer_ade20k_swin_tiny](https://huggingface.co/shi-labs/oneformer_ade20k_swin_tiny) architecture34    trained on [ADE20k-150](https://huggingface.co/datasets/scene_parse_150).35 36    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the37    documentation from [`PretrainedConfig`] for more information.38 39    Args:40        backbone_config (`PretrainedConfig`, *optional*, defaults to `SwinConfig`):41            The configuration of the backbone model.42        backbone (`str`, *optional*):43            Name of backbone to use when `backbone_config` is `None`. If `use_pretrained_backbone` is `True`, this44            will load the corresponding pretrained weights from the timm or transformers library. If `use_pretrained_backbone`45            is `False`, this loads the backbone's config and uses that to initialize the backbone with random weights.46        use_pretrained_backbone (`bool`, *optional*, defaults to `False`):47            Whether to use pretrained weights for the backbone.48        use_timm_backbone (`bool`, *optional*, defaults to `False`):49            Whether to load `backbone` from the timm library. If `False`, the backbone is loaded from the transformers50            library.51        backbone_kwargs (`dict`, *optional*):52            Keyword arguments to be passed to AutoBackbone when loading from a checkpoint53            e.g. `{'out_indices': (0, 1, 2, 3)}`. Cannot be specified if `backbone_config` is set.54        ignore_value (`int`, *optional*, defaults to 255):55            Values to be ignored in GT label while calculating loss.56        num_queries (`int`, *optional*, defaults to 150):57            Number of object queries.58        no_object_weight (`float`, *optional*, defaults to 0.1):59            Weight for no-object class predictions.60        class_weight (`float`, *optional*, defaults to 2.0):61            Weight for Classification CE loss.62        mask_weight (`float`, *optional*, defaults to 5.0):63            Weight for binary CE loss.64        dice_weight (`float`, *optional*, defaults to 5.0):65            Weight for dice loss.66        contrastive_weight (`float`, *optional*, defaults to 0.5):67            Weight for contrastive loss.68        contrastive_temperature (`float`, *optional*, defaults to 0.07):69            Initial value for scaling the contrastive logits.70        train_num_points (`int`, *optional*, defaults to 12544):71            Number of points to sample while calculating losses on mask predictions.72        oversample_ratio (`float`, *optional*, defaults to 3.0):73            Ratio to decide how many points to oversample.74        importance_sample_ratio (`float`, *optional*, defaults to 0.75):75            Ratio of points that are sampled via importance sampling.76        init_std (`float`, *optional*, defaults to 0.02):77            Standard deviation for normal initialization.78        init_xavier_std (`float`, *optional*, defaults to 1.0):79            Standard deviation for xavier uniform initialization.80        layer_norm_eps (`float`, *optional*, defaults to 1e-05):81            Epsilon for layer normalization.82        is_training (`bool`, *optional*, defaults to `False`):83            Whether to run in training or inference mode.84        use_auxiliary_loss (`bool`, *optional*, defaults to `True`):85            Whether to calculate loss using intermediate predictions from transformer decoder.86        output_auxiliary_logits (`bool`, *optional*, defaults to `True`):87            Whether to return intermediate predictions from transformer decoder.88        strides (`list`, *optional*, defaults to `[4, 8, 16, 32]`):89            List containing the strides for feature maps in the encoder.90        task_seq_len (`int`, *optional*, defaults to 77):91            Sequence length for tokenizing text list input.92        text_encoder_width (`int`, *optional*, defaults to 256):93            Hidden size for text encoder.94        text_encoder_context_length (`int`, *optional*, defaults to 77):95            Input sequence length for text encoder.96        text_encoder_num_layers (`int`, *optional*, defaults to 6):97            Number of layers for transformer in text encoder.98        text_encoder_vocab_size (`int`, *optional*, defaults to 49408):99            Vocabulary size for tokenizer.100        text_encoder_proj_layers (`int`, *optional*, defaults to 2):101            Number of layers in MLP for project text queries.102        text_encoder_n_ctx (`int`, *optional*, defaults to 16):103            Number of learnable text context queries.104        conv_dim (`int`, *optional*, defaults to 256):105            Feature map dimension to map outputs from the backbone.106        mask_dim (`int`, *optional*, defaults to 256):107            Dimension for feature maps in pixel decoder.108        hidden_dim (`int`, *optional*, defaults to 256):109            Dimension for hidden states in transformer decoder.110        encoder_feedforward_dim (`int`, *optional*, defaults to 1024):111            Dimension for FFN layer in pixel decoder.112        norm (`str`, *optional*, defaults to `"GN"`):113            Type of normalization.114        encoder_layers (`int`, *optional*, defaults to 6):115            Number of layers in pixel decoder.116        decoder_layers (`int`, *optional*, defaults to 10):117            Number of layers in transformer decoder.118        use_task_norm (`bool`, *optional*, defaults to `True`):119            Whether to normalize the task token.120        num_attention_heads (`int`, *optional*, defaults to 8):121            Number of attention heads in transformer layers in the pixel and transformer decoders.122        dropout (`float`, *optional*, defaults to 0.1):123            Dropout probability for pixel and transformer decoders.124        dim_feedforward (`int`, *optional*, defaults to 2048):125            Dimension for FFN layer in transformer decoder.126        pre_norm (`bool`, *optional*, defaults to `False`):127            Whether to normalize hidden states before attention layers in transformer decoder.128        enforce_input_proj (`bool`, *optional*, defaults to `False`):129            Whether to project hidden states in transformer decoder.130        query_dec_layers (`int`, *optional*, defaults to 2):131            Number of layers in query transformer.132        common_stride (`int`, *optional*, defaults to 4):133            Common stride used for features in pixel decoder.134 135    Examples:136    ```python137    >>> from transformers import OneFormerConfig, OneFormerModel138 139    >>> # Initializing a OneFormer shi-labs/oneformer_ade20k_swin_tiny configuration140    >>> configuration = OneFormerConfig()141    >>> # Initializing a model (with random weights) from the shi-labs/oneformer_ade20k_swin_tiny style configuration142    >>> model = OneFormerModel(configuration)143    >>> # Accessing the model configuration144    >>> configuration = model.config145    ```146    """147 148    model_type = "oneformer"149    attribute_map = {"hidden_size": "hidden_dim"}150 151    def __init__(152        self,153        backbone_config: Optional[dict] = None,154        backbone: Optional[str] = None,155        use_pretrained_backbone: bool = False,156        use_timm_backbone: bool = False,157        backbone_kwargs: Optional[dict] = None,158        ignore_value: int = 255,159        num_queries: int = 150,160        no_object_weight: int = 0.1,161        class_weight: float = 2.0,162        mask_weight: float = 5.0,163        dice_weight: float = 5.0,164        contrastive_weight: float = 0.5,165        contrastive_temperature: float = 0.07,166        train_num_points: int = 12544,167        oversample_ratio: float = 3.0,168        importance_sample_ratio: float = 0.75,169        init_std: float = 0.02,170        init_xavier_std: float = 1.0,171        layer_norm_eps: float = 1e-05,172        is_training: bool = False,173        use_auxiliary_loss: bool = True,174        output_auxiliary_logits: bool = True,175        strides: Optional[list] = [4, 8, 16, 32],176        task_seq_len: int = 77,177        text_encoder_width: int = 256,178        text_encoder_context_length: int = 77,179        text_encoder_num_layers: int = 6,180        text_encoder_vocab_size: int = 49408,181        text_encoder_proj_layers: int = 2,182        text_encoder_n_ctx: int = 16,183        conv_dim: int = 256,184        mask_dim: int = 256,185        hidden_dim: int = 256,186        encoder_feedforward_dim: int = 1024,187        norm: str = "GN",188        encoder_layers: int = 6,189        decoder_layers: int = 10,190        use_task_norm: bool = True,191        num_attention_heads: int = 8,192        dropout: float = 0.1,193        dim_feedforward: int = 2048,194        pre_norm: bool = False,195        enforce_input_proj: bool = False,196        query_dec_layers: int = 2,197        common_stride: int = 4,198        **kwargs,199    ):200        if backbone_config is None and backbone is None:201            logger.info("`backbone_config` is unset. Initializing the config with the default `Swin` backbone.")202            backbone_config = CONFIG_MAPPING["swin"](203                image_size=224,204                num_channels=3,205                patch_size=4,206                embed_dim=96,207                depths=[2, 2, 6, 2],208                num_heads=[3, 6, 12, 24],209                window_size=7,210                drop_path_rate=0.3,211                use_absolute_embeddings=False,212                out_features=["stage1", "stage2", "stage3", "stage4"],213            )214        elif isinstance(backbone_config, dict):215            backbone_model_type = backbone_config.get("model_type")216            config_class = CONFIG_MAPPING[backbone_model_type]217            backbone_config = config_class.from_dict(backbone_config)218 219        verify_backbone_config_arguments(220            use_timm_backbone=use_timm_backbone,221            use_pretrained_backbone=use_pretrained_backbone,222            backbone=backbone,223            backbone_config=backbone_config,224            backbone_kwargs=backbone_kwargs,225        )226 227        self.backbone_config = backbone_config228        self.backbone = backbone229        self.use_pretrained_backbone = use_pretrained_backbone230        self.use_timm_backbone = use_timm_backbone231        self.backbone_kwargs = backbone_kwargs232        self.ignore_value = ignore_value233        self.num_queries = num_queries234        self.no_object_weight = no_object_weight235        self.class_weight = class_weight236        self.mask_weight = mask_weight237        self.dice_weight = dice_weight238        self.contrastive_weight = contrastive_weight239        self.contrastive_temperature = contrastive_temperature240        self.train_num_points = train_num_points241        self.oversample_ratio = oversample_ratio242        self.importance_sample_ratio = importance_sample_ratio243        self.init_std = init_std244        self.init_xavier_std = init_xavier_std245        self.layer_norm_eps = layer_norm_eps246        self.is_training = is_training247        self.use_auxiliary_loss = use_auxiliary_loss248        self.output_auxiliary_logits = output_auxiliary_logits249        self.strides = strides250        self.task_seq_len = task_seq_len251        self.text_encoder_width = text_encoder_width252        self.text_encoder_context_length = text_encoder_context_length253        self.text_encoder_num_layers = text_encoder_num_layers254        self.text_encoder_vocab_size = text_encoder_vocab_size255        self.text_encoder_proj_layers = text_encoder_proj_layers256        self.text_encoder_n_ctx = text_encoder_n_ctx257        self.conv_dim = conv_dim258        self.mask_dim = mask_dim259        self.hidden_dim = hidden_dim260        self.encoder_feedforward_dim = encoder_feedforward_dim261        self.norm = norm262        self.encoder_layers = encoder_layers263        self.decoder_layers = decoder_layers264        self.use_task_norm = use_task_norm265        self.num_attention_heads = num_attention_heads266        self.dropout = dropout267        self.dim_feedforward = dim_feedforward268        self.pre_norm = pre_norm269        self.enforce_input_proj = enforce_input_proj270        self.query_dec_layers = query_dec_layers271        self.common_stride = common_stride272        self.num_hidden_layers = decoder_layers273 274        super().__init__(**kwargs)275 276    @property277    def sub_configs(self):278        return (279            {"backbone_config": type(self.backbone_config)}280            if getattr(self, "backbone_config", None) is not None281            else {}282        )283 284 285__all__ = ["OneFormerConfig"]286