Aluode/PerceptionLabPortable
0
1# coding=utf-82# Copyright 2022 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"""CLIPSeg model configuration"""16 17from ...configuration_utils import PretrainedConfig18from ...utils import logging19 20 21logger = logging.get_logger(__name__)22 23 24class CLIPSegTextConfig(PretrainedConfig):25 r"""26 This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an27 CLIPSeg model according to the specified arguments, defining the model architecture. Instantiating a configuration28 with the defaults will yield a similar configuration to that of the CLIPSeg29 [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.30 31 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the32 documentation from [`PretrainedConfig`] for more information.33 34 Args:35 vocab_size (`int`, *optional*, defaults to 49408):36 Vocabulary size of the CLIPSeg text model. Defines the number of different tokens that can be represented37 by the `inputs_ids` passed when calling [`CLIPSegModel`].38 hidden_size (`int`, *optional*, defaults to 512):39 Dimensionality of the encoder layers and the pooler layer.40 intermediate_size (`int`, *optional*, defaults to 2048):41 Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.42 num_hidden_layers (`int`, *optional*, defaults to 12):43 Number of hidden layers in the Transformer encoder.44 num_attention_heads (`int`, *optional*, defaults to 8):45 Number of attention heads for each attention layer in the Transformer encoder.46 max_position_embeddings (`int`, *optional*, defaults to 77):47 The maximum sequence length that this model might ever be used with. Typically set this to something large48 just in case (e.g., 512 or 1024 or 2048).49 hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):50 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,51 `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.52 layer_norm_eps (`float`, *optional*, defaults to 1e-05):53 The epsilon used by the layer normalization layers.54 attention_dropout (`float`, *optional*, defaults to 0.0):55 The dropout ratio for the attention probabilities.56 initializer_range (`float`, *optional*, defaults to 0.02):57 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.58 initializer_factor (`float`, *optional*, defaults to 1.0):59 A factor for initializing all weight matrices (should be kept to 1, used internally for initialization60 testing).61 pad_token_id (`int`, *optional*, defaults to 1):62 Padding token id.63 bos_token_id (`int`, *optional*, defaults to 49406):64 Beginning of stream token id.65 eos_token_id (`int`, *optional*, defaults to 49407):66 End of stream token id.67 68 Example:69 70 ```python71 >>> from transformers import CLIPSegTextConfig, CLIPSegTextModel72 73 >>> # Initializing a CLIPSegTextConfig with CIDAS/clipseg-rd64 style configuration74 >>> configuration = CLIPSegTextConfig()75 76 >>> # Initializing a CLIPSegTextModel (with random weights) from the CIDAS/clipseg-rd64 style configuration77 >>> model = CLIPSegTextModel(configuration)78 79 >>> # Accessing the model configuration80 >>> configuration = model.config81 ```"""82 83 model_type = "clipseg_text_model"84 base_config_key = "text_config"85 86 def __init__(87 self,88 vocab_size=49408,89 hidden_size=512,90 intermediate_size=2048,91 num_hidden_layers=12,92 num_attention_heads=8,93 max_position_embeddings=77,94 hidden_act="quick_gelu",95 layer_norm_eps=1e-5,96 attention_dropout=0.0,97 initializer_range=0.02,98 initializer_factor=1.0,99 pad_token_id=1,100 bos_token_id=49406,101 eos_token_id=49407,102 **kwargs,103 ):104 super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)105 106 self.vocab_size = vocab_size107 self.hidden_size = hidden_size108 self.intermediate_size = intermediate_size109 self.num_hidden_layers = num_hidden_layers110 self.num_attention_heads = num_attention_heads111 self.max_position_embeddings = max_position_embeddings112 self.layer_norm_eps = layer_norm_eps113 self.hidden_act = hidden_act114 self.initializer_range = initializer_range115 self.initializer_factor = initializer_factor116 self.attention_dropout = attention_dropout117 118 119class CLIPSegVisionConfig(PretrainedConfig):120 r"""121 This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an122 CLIPSeg model according to the specified arguments, defining the model architecture. Instantiating a configuration123 with the defaults will yield a similar configuration to that of the CLIPSeg124 [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.125 126 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the127 documentation from [`PretrainedConfig`] for more information.128 129 Args:130 hidden_size (`int`, *optional*, defaults to 768):131 Dimensionality of the encoder layers and the pooler layer.132 intermediate_size (`int`, *optional*, defaults to 3072):133 Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.134 num_hidden_layers (`int`, *optional*, defaults to 12):135 Number of hidden layers in the Transformer encoder.136 num_attention_heads (`int`, *optional*, defaults to 12):137 Number of attention heads for each attention layer in the Transformer encoder.138 num_channels (`int`, *optional*, defaults to 3):139 The number of input channels.140 image_size (`int`, *optional*, defaults to 224):141 The size (resolution) of each image.142 patch_size (`int`, *optional*, defaults to 32):143 The size (resolution) of each patch.144 hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):145 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,146 `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.147 layer_norm_eps (`float`, *optional*, defaults to 1e-05):148 The epsilon used by the layer normalization layers.149 attention_dropout (`float`, *optional*, defaults to 0.0):150 The dropout ratio for the attention probabilities.151 initializer_range (`float`, *optional*, defaults to 0.02):152 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.153 initializer_factor (`float`, *optional*, defaults to 1.0):154 A factor for initializing all weight matrices (should be kept to 1, used internally for initialization155 testing).156 157 Example:158 159 ```python160 >>> from transformers import CLIPSegVisionConfig, CLIPSegVisionModel161 162 >>> # Initializing a CLIPSegVisionConfig with CIDAS/clipseg-rd64 style configuration163 >>> configuration = CLIPSegVisionConfig()164 165 >>> # Initializing a CLIPSegVisionModel (with random weights) from the CIDAS/clipseg-rd64 style configuration166 >>> model = CLIPSegVisionModel(configuration)167 168 >>> # Accessing the model configuration169 >>> configuration = model.config170 ```"""171 172 model_type = "clipseg_vision_model"173 base_config_key = "vision_config"174 175 def __init__(176 self,177 hidden_size=768,178 intermediate_size=3072,179 num_hidden_layers=12,180 num_attention_heads=12,181 num_channels=3,182 image_size=224,183 patch_size=32,184 hidden_act="quick_gelu",185 layer_norm_eps=1e-5,186 attention_dropout=0.0,187 initializer_range=0.02,188 initializer_factor=1.0,189 **kwargs,190 ):191 super().__init__(**kwargs)192 193 self.hidden_size = hidden_size194 self.intermediate_size = intermediate_size195 self.num_hidden_layers = num_hidden_layers196 self.num_attention_heads = num_attention_heads197 self.num_channels = num_channels198 self.patch_size = patch_size199 self.image_size = image_size200 self.initializer_range = initializer_range201 self.initializer_factor = initializer_factor202 self.attention_dropout = attention_dropout203 self.layer_norm_eps = layer_norm_eps204 self.hidden_act = hidden_act205 206 207class CLIPSegConfig(PretrainedConfig):208 r"""209 [`CLIPSegConfig`] is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to210 instantiate a CLIPSeg model according to the specified arguments, defining the text model and vision model configs.211 Instantiating a configuration with the defaults will yield a similar configuration to that of the CLIPSeg212 [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.213 214 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the215 documentation from [`PretrainedConfig`] for more information.216 217 Args:218 text_config (`dict`, *optional*):219 Dictionary of configuration options used to initialize [`CLIPSegTextConfig`].220 vision_config (`dict`, *optional*):221 Dictionary of configuration options used to initialize [`CLIPSegVisionConfig`].222 projection_dim (`int`, *optional*, defaults to 512):223 Dimensionality of text and vision projection layers.224 logit_scale_init_value (`float`, *optional*, defaults to 2.6592):225 The initial value of the *logit_scale* parameter. Default is used as per the original CLIPSeg implementation.226 extract_layers (`list[int]`, *optional*, defaults to `[3, 6, 9]`):227 Layers to extract when forwarding the query image through the frozen visual backbone of CLIP.228 reduce_dim (`int`, *optional*, defaults to 64):229 Dimensionality to reduce the CLIP vision embedding.230 decoder_num_attention_heads (`int`, *optional*, defaults to 4):231 Number of attention heads in the decoder of CLIPSeg.232 decoder_attention_dropout (`float`, *optional*, defaults to 0.0):233 The dropout ratio for the attention probabilities.234 decoder_hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):235 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,236 `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.237 decoder_intermediate_size (`int`, *optional*, defaults to 2048):238 Dimensionality of the "intermediate" (i.e., feed-forward) layers in the Transformer decoder.239 conditional_layer (`int`, *optional*, defaults to 0):240 The layer to use of the Transformer encoder whose activations will be combined with the condition241 embeddings using FiLM (Feature-wise Linear Modulation). If 0, the last layer is used.242 use_complex_transposed_convolution (`bool`, *optional*, defaults to `False`):243 Whether to use a more complex transposed convolution in the decoder, enabling more fine-grained244 segmentation.245 kwargs (*optional*):246 Dictionary of keyword arguments.247 248 Example:249 250 ```python251 >>> from transformers import CLIPSegConfig, CLIPSegModel252 253 >>> # Initializing a CLIPSegConfig with CIDAS/clipseg-rd64 style configuration254 >>> configuration = CLIPSegConfig()255 256 >>> # Initializing a CLIPSegModel (with random weights) from the CIDAS/clipseg-rd64 style configuration257 >>> model = CLIPSegModel(configuration)258 259 >>> # Accessing the model configuration260 >>> configuration = model.config261 262 >>> # We can also initialize a CLIPSegConfig from a CLIPSegTextConfig and a CLIPSegVisionConfig263 264 >>> # Initializing a CLIPSegText and CLIPSegVision configuration265 >>> config_text = CLIPSegTextConfig()266 >>> config_vision = CLIPSegVisionConfig()267 268 >>> config = CLIPSegConfig.from_text_vision_configs(config_text, config_vision)269 ```"""270 271 model_type = "clipseg"272 sub_configs = {"text_config": CLIPSegTextConfig, "vision_config": CLIPSegVisionConfig}273 274 def __init__(275 self,276 text_config=None,277 vision_config=None,278 projection_dim=512,279 logit_scale_init_value=2.6592,280 extract_layers=[3, 6, 9],281 reduce_dim=64,282 decoder_num_attention_heads=4,283 decoder_attention_dropout=0.0,284 decoder_hidden_act="quick_gelu",285 decoder_intermediate_size=2048,286 conditional_layer=0,287 use_complex_transposed_convolution=False,288 **kwargs,289 ):290 # If `_config_dict` exist, we use them for the backward compatibility.291 # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot292 # of confusion!).293 text_config_dict = kwargs.pop("text_config_dict", None)294 vision_config_dict = kwargs.pop("vision_config_dict", None)295 296 super().__init__(**kwargs)297 298 # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in299 # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most300 # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.301 if text_config_dict is not None:302 if text_config is None:303 text_config = {}304 305 # This is the complete result when using `text_config_dict`.306 _text_config_dict = CLIPSegTextConfig(**text_config_dict).to_dict()307 308 # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.309 for key, value in _text_config_dict.items():310 if key in text_config and value != text_config[key] and key != "transformers_version":311 # If specified in `text_config_dict`312 if key in text_config_dict:313 message = (314 f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "315 f'The value `text_config_dict["{key}"]` will be used instead.'316 )317 # If inferred from default argument values (just to be super careful)318 else:319 message = (320 f"`text_config_dict` is provided which will be used to initialize `CLIPSegTextConfig`. The "321 f'value `text_config["{key}"]` will be overridden.'322 )323 logger.info(message)324 325 # Update all values in `text_config` with the ones in `_text_config_dict`.326 text_config.update(_text_config_dict)327 328 if vision_config_dict is not None:329 if vision_config is None:330 vision_config = {}331 332 # This is the complete result when using `vision_config_dict`.333 _vision_config_dict = CLIPSegVisionConfig(**vision_config_dict).to_dict()334 # convert keys to string instead of integer335 if "id2label" in _vision_config_dict:336 _vision_config_dict["id2label"] = {337 str(key): value for key, value in _vision_config_dict["id2label"].items()338 }339 340 # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.341 for key, value in _vision_config_dict.items():342 if key in vision_config and value != vision_config[key] and key != "transformers_version":343 # If specified in `vision_config_dict`344 if key in vision_config_dict:345 message = (346 f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "347 f'values. The value `vision_config_dict["{key}"]` will be used instead.'348 )349 # If inferred from default argument values (just to be super careful)350 else:351 message = (352 f"`vision_config_dict` is provided which will be used to initialize `CLIPSegVisionConfig`. "353 f'The value `vision_config["{key}"]` will be overridden.'354 )355 logger.info(message)356 357 # Update all values in `vision_config` with the ones in `_vision_config_dict`.358 vision_config.update(_vision_config_dict)359 360 if text_config is None:361 text_config = {}362 logger.info("`text_config` is `None`. Initializing the `CLIPSegTextConfig` with default values.")363 364 if vision_config is None:365 vision_config = {}366 logger.info("`vision_config` is `None`. initializing the `CLIPSegVisionConfig` with default values.")367 368 self.text_config = CLIPSegTextConfig(**text_config)369 self.vision_config = CLIPSegVisionConfig(**vision_config)370 371 self.projection_dim = projection_dim372 self.logit_scale_init_value = logit_scale_init_value373 self.extract_layers = extract_layers374 self.reduce_dim = reduce_dim375 self.decoder_num_attention_heads = decoder_num_attention_heads376 self.decoder_attention_dropout = decoder_attention_dropout377 self.decoder_hidden_act = decoder_hidden_act378 self.decoder_intermediate_size = decoder_intermediate_size379 self.conditional_layer = conditional_layer380 self.initializer_factor = 1.0381 self.use_complex_transposed_convolution = use_complex_transposed_convolution382 383 384__all__ = ["CLIPSegConfig", "CLIPSegTextConfig", "CLIPSegVisionConfig"]385