Aluode/PerceptionLabPortable
0
1# coding=utf-82# Copyright 2022 The OFA-Sys Team Authors and The HuggingFace 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"""Chinese-CLIP model configuration"""16 17from collections import OrderedDict18from collections.abc import Mapping19from typing import TYPE_CHECKING, Any, Optional20 21 22if TYPE_CHECKING:23 from ...processing_utils import ProcessorMixin24 from ...utils import TensorType25 26from ...configuration_utils import PretrainedConfig27from ...onnx import OnnxConfig28from ...utils import logging29 30 31logger = logging.get_logger(__name__)32 33 34class ChineseCLIPTextConfig(PretrainedConfig):35 r"""36 This is the configuration class to store the configuration of a [`ChineseCLIPModel`]. It is used to instantiate a37 Chinese CLIP model according to the specified arguments, defining the model architecture. Instantiating a38 configuration with the defaults will yield a similar configuration to that of the Chinese CLIP39 [OFA-Sys/chinese-clip-vit-base-patch16](https:40 //huggingface.co/OFA-Sys/chinese-clip-vit-base-patch16) architecture.41 42 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the43 documentation from [`PretrainedConfig`] for more information.44 45 46 Args:47 vocab_size (`int`, *optional*, defaults to 30522):48 Vocabulary size of the CHINESE_CLIP model. Defines the number of different tokens that can be represented49 by the `inputs_ids` passed when calling [`ChineseCLIPModel`].50 hidden_size (`int`, *optional*, defaults to 768):51 Dimensionality of the encoder layers and the pooler layer.52 num_hidden_layers (`int`, *optional*, defaults to 12):53 Number of hidden layers in the Transformer encoder.54 num_attention_heads (`int`, *optional*, defaults to 12):55 Number of attention heads for each attention layer in the Transformer encoder.56 intermediate_size (`int`, *optional*, defaults to 3072):57 Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.58 hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):59 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,60 `"relu"`, `"silu"` and `"gelu_new"` are supported.61 hidden_dropout_prob (`float`, *optional*, defaults to 0.1):62 The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.63 attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):64 The dropout ratio for the attention probabilities.65 max_position_embeddings (`int`, *optional*, defaults to 512):66 The maximum sequence length that this model might ever be used with. Typically set this to something large67 just in case (e.g., 512 or 1024 or 2048).68 type_vocab_size (`int`, *optional*, defaults to 2):69 The vocabulary size of the `token_type_ids` passed when calling [`ChineseCLIPModel`].70 initializer_range (`float`, *optional*, defaults to 0.02):71 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.72 initializer_factor (`float`, *optional*, defaults to 1.0):73 A factor for initializing all weight matrices (should be kept to 1, used internally for initialization74 testing).75 layer_norm_eps (`float`, *optional*, defaults to 1e-12):76 The epsilon used by the layer normalization layers.77 pad_token_id (`int`, *optional*, defaults to 0):78 Padding token id.79 position_embedding_type (`str`, *optional*, defaults to `"absolute"`):80 Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For81 positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to82 [Self-Attention with Relative Position Representations (Shaw et al.)](https://huggingface.co/papers/1803.02155).83 For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models84 with Better Relative Position Embeddings (Huang et al.)](https://huggingface.co/papers/2009.13658).85 use_cache (`bool`, *optional*, defaults to `True`):86 Whether or not the model should return the last key/values attentions (not used by all models). Only87 relevant if `config.is_decoder=True`.88 89 Example:90 91 ```python92 >>> from transformers import ChineseCLIPTextConfig, ChineseCLIPTextModel93 94 >>> # Initializing a ChineseCLIPTextConfig with OFA-Sys/chinese-clip-vit-base-patch16 style configuration95 >>> configuration = ChineseCLIPTextConfig()96 97 >>> # Initializing a ChineseCLIPTextModel (with random weights) from the OFA-Sys/chinese-clip-vit-base-patch16 style configuration98 >>> model = ChineseCLIPTextModel(configuration)99 100 >>> # Accessing the model configuration101 >>> configuration = model.config102 ```"""103 104 model_type = "chinese_clip_text_model"105 base_config_key = "text_config"106 107 def __init__(108 self,109 vocab_size=30522,110 hidden_size=768,111 num_hidden_layers=12,112 num_attention_heads=12,113 intermediate_size=3072,114 hidden_act="gelu",115 hidden_dropout_prob=0.1,116 attention_probs_dropout_prob=0.1,117 max_position_embeddings=512,118 type_vocab_size=2,119 initializer_range=0.02,120 initializer_factor=1.0,121 layer_norm_eps=1e-12,122 pad_token_id=0,123 position_embedding_type="absolute",124 use_cache=True,125 **kwargs,126 ):127 super().__init__(pad_token_id=pad_token_id, **kwargs)128 129 self.vocab_size = vocab_size130 self.hidden_size = hidden_size131 self.num_hidden_layers = num_hidden_layers132 self.num_attention_heads = num_attention_heads133 self.hidden_act = hidden_act134 self.intermediate_size = intermediate_size135 self.hidden_dropout_prob = hidden_dropout_prob136 self.attention_probs_dropout_prob = attention_probs_dropout_prob137 self.max_position_embeddings = max_position_embeddings138 self.type_vocab_size = type_vocab_size139 self.initializer_range = initializer_range140 self.initializer_factor = initializer_factor141 self.layer_norm_eps = layer_norm_eps142 self.position_embedding_type = position_embedding_type143 self.use_cache = use_cache144 145 146class ChineseCLIPVisionConfig(PretrainedConfig):147 r"""148 This is the configuration class to store the configuration of a [`ChineseCLIPModel`]. It is used to instantiate an149 ChineseCLIP model according to the specified arguments, defining the model architecture. Instantiating a150 configuration with the defaults will yield a similar configuration to that of the ChineseCLIP151 [OFA-Sys/chinese-clip-vit-base-patch16](https://huggingface.co/OFA-Sys/chinese-clip-vit-base-patch16) architecture.152 153 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the154 documentation from [`PretrainedConfig`] for more information.155 156 157 Args:158 hidden_size (`int`, *optional*, defaults to 768):159 Dimensionality of the encoder layers and the pooler layer.160 intermediate_size (`int`, *optional*, defaults to 3072):161 Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.162 projection_dim (`int`, *optional*, defaults to 512):163 Dimensionality of text and vision projection layers.164 num_hidden_layers (`int`, *optional*, defaults to 12):165 Number of hidden layers in the Transformer encoder.166 num_attention_heads (`int`, *optional*, defaults to 12):167 Number of attention heads for each attention layer in the Transformer encoder.168 num_channels (`int`, *optional*, defaults to 3):169 The number of input channels.170 image_size (`int`, *optional*, defaults to 224):171 The size (resolution) of each image.172 patch_size (`int`, *optional*, defaults to 32):173 The size (resolution) of each patch.174 hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):175 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,176 `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.177 layer_norm_eps (`float`, *optional*, defaults to 1e-05):178 The epsilon used by the layer normalization layers.179 attention_dropout (`float`, *optional*, defaults to 0.0):180 The dropout ratio for the attention probabilities.181 initializer_range (`float`, *optional*, defaults to 0.02):182 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.183 initializer_factor (`float`, *optional*, defaults to 1.0):184 A factor for initializing all weight matrices (should be kept to 1, used internally for initialization185 testing).186 Example:187 ```python188 >>> from transformers import ChineseCLIPVisionConfig, ChineseCLIPVisionModel189 190 >>> # Initializing a ChineseCLIPVisionConfig with OFA-Sys/chinese-clip-vit-base-patch16 style configuration191 >>> configuration = ChineseCLIPVisionConfig()192 193 >>> # Initializing a ChineseCLIPVisionModel (with random weights) from the OFA-Sys/chinese-clip-vit-base-patch16 style configuration194 >>> model = ChineseCLIPVisionModel(configuration)195 196 >>> # Accessing the model configuration197 >>> configuration = model.config198 ```"""199 200 model_type = "chinese_clip_vision_model"201 base_config_key = "vision_config"202 203 def __init__(204 self,205 hidden_size=768,206 intermediate_size=3072,207 projection_dim=512,208 num_hidden_layers=12,209 num_attention_heads=12,210 num_channels=3,211 image_size=224,212 patch_size=32,213 hidden_act="quick_gelu",214 layer_norm_eps=1e-5,215 attention_dropout=0.0,216 initializer_range=0.02,217 initializer_factor=1.0,218 **kwargs,219 ):220 super().__init__(**kwargs)221 222 self.hidden_size = hidden_size223 self.intermediate_size = intermediate_size224 self.projection_dim = projection_dim225 self.num_hidden_layers = num_hidden_layers226 self.num_attention_heads = num_attention_heads227 self.num_channels = num_channels228 self.patch_size = patch_size229 self.image_size = image_size230 self.initializer_range = initializer_range231 self.initializer_factor = initializer_factor232 self.attention_dropout = attention_dropout233 self.layer_norm_eps = layer_norm_eps234 self.hidden_act = hidden_act235 236 237class ChineseCLIPConfig(PretrainedConfig):238 r"""239 [`ChineseCLIPConfig`] is the configuration class to store the configuration of a [`ChineseCLIPModel`]. It is used240 to instantiate Chinese-CLIP model according to the specified arguments, defining the text model and vision model241 configs. Instantiating a configuration with the defaults will yield a similar configuration to that of the242 Chinese-CLIP [OFA-Sys/chinese-clip-vit-base-patch16](https://huggingface.co/OFA-Sys/chinese-clip-vit-base-patch16)243 architecture.244 245 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the246 documentation from [`PretrainedConfig`] for more information.247 248 Args:249 text_config (`dict`, *optional*):250 Dictionary of configuration options used to initialize [`ChineseCLIPTextConfig`].251 vision_config (`dict`, *optional*):252 Dictionary of configuration options used to initialize [`ChineseCLIPVisionConfig`].253 projection_dim (`int`, *optional*, defaults to 512):254 Dimensionality of text and vision projection layers.255 logit_scale_init_value (`float`, *optional*, defaults to 2.6592):256 The initial value of the *logit_scale* parameter. Default is used as per the original ChineseCLIP257 implementation.258 kwargs (*optional*):259 Dictionary of keyword arguments.260 261 Example:262 263 ```python264 >>> from transformers import ChineseCLIPConfig, ChineseCLIPModel265 266 >>> # Initializing a ChineseCLIPConfig with OFA-Sys/chinese-clip-vit-base-patch16 style configuration267 >>> configuration = ChineseCLIPConfig()268 269 >>> # Initializing a ChineseCLIPModel (with random weights) from the OFA-Sys/chinese-clip-vit-base-patch16 style configuration270 >>> model = ChineseCLIPModel(configuration)271 272 >>> # Accessing the model configuration273 >>> configuration = model.config274 275 >>> # We can also initialize a ChineseCLIPConfig from a ChineseCLIPTextConfig and a ChineseCLIPVisionConfig276 277 >>> # Initializing a ChineseCLIPTextConfig and ChineseCLIPVisionConfig configuration278 >>> config_text = ChineseCLIPTextConfig()279 >>> config_vision = ChineseCLIPVisionConfig()280 281 >>> config = ChineseCLIPConfig.from_text_vision_configs(config_text, config_vision)282 ```"""283 284 model_type = "chinese_clip"285 sub_configs = {"text_config": ChineseCLIPTextConfig, "vision_config": ChineseCLIPVisionConfig}286 287 def __init__(288 self, text_config=None, vision_config=None, projection_dim=512, logit_scale_init_value=2.6592, **kwargs289 ):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 = ChineseCLIPTextConfig(**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 `ChineseCLIPTextConfig`. "321 f'The 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 = ChineseCLIPVisionConfig(**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 "353 f'`ChineseCLIPVisionConfig`. 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 `ChineseCLIPTextConfig` with default values.")363 364 if vision_config is None:365 vision_config = {}366 logger.info("`vision_config` is `None`. initializing the `ChineseCLIPVisionConfig` with default values.")367 368 self.text_config = ChineseCLIPTextConfig(**text_config)369 self.vision_config = ChineseCLIPVisionConfig(**vision_config)370 371 self.projection_dim = projection_dim372 self.logit_scale_init_value = logit_scale_init_value373 self.initializer_factor = 1.0374 self.initializer_range = 0.02375 376 377class ChineseCLIPOnnxConfig(OnnxConfig):378 @property379 def inputs(self) -> Mapping[str, Mapping[int, str]]:380 return OrderedDict(381 [382 ("input_ids", {0: "batch", 1: "sequence"}),383 ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),384 ("attention_mask", {0: "batch", 1: "sequence"}),385 ]386 )387 388 @property389 def outputs(self) -> Mapping[str, Mapping[int, str]]:390 return OrderedDict(391 [392 ("logits_per_image", {0: "batch"}),393 ("logits_per_text", {0: "batch"}),394 ("text_embeds", {0: "batch"}),395 ("image_embeds", {0: "batch"}),396 ]397 )398 399 @property400 def atol_for_validation(self) -> float:401 return 1e-4402 403 def generate_dummy_inputs(404 self,405 processor: "ProcessorMixin",406 batch_size: int = -1,407 seq_length: int = -1,408 framework: Optional["TensorType"] = None,409 ) -> Mapping[str, Any]:410 text_input_dict = super().generate_dummy_inputs(411 processor.tokenizer, batch_size=batch_size, seq_length=seq_length, framework=framework412 )413 image_input_dict = super().generate_dummy_inputs(414 processor.image_processor, batch_size=batch_size, framework=framework415 )416 return {**text_input_dict, **image_input_dict}417 418 @property419 def default_onnx_opset(self) -> int:420 return 14421 422 423__all__ = ["ChineseCLIPConfig", "ChineseCLIPOnnxConfig", "ChineseCLIPTextConfig", "ChineseCLIPVisionConfig"]424