Aluode/PerceptionLabPortable
0
1# coding=utf-82# Copyright 2021 The Fairseq Authors 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"""Wav2Vec2 model configuration"""16 17import functools18import operator19 20from ...configuration_utils import PretrainedConfig21from ...utils import logging22 23 24logger = logging.get_logger(__name__)25 26 27class Wav2Vec2Config(PretrainedConfig):28 r"""29 This is the configuration class to store the configuration of a [`Wav2Vec2Model`]. It is used to instantiate an30 Wav2Vec2 model according to the specified arguments, defining the model architecture. Instantiating a configuration31 with the defaults will yield a similar configuration to that of the Wav2Vec232 [facebook/wav2vec2-base-960h](https://huggingface.co/facebook/wav2vec2-base-960h) architecture.33 34 Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the35 documentation from [`PretrainedConfig`] for more information.36 37 38 Args:39 vocab_size (`int`, *optional*, defaults to 32):40 Vocabulary size of the Wav2Vec2 model. Defines the number of different tokens that can be represented by41 the `inputs_ids` passed when calling [`Wav2Vec2Model`] or [`TFWav2Vec2Model`]. Vocabulary size of the42 model. Defines the different tokens that can be represented by the *inputs_ids* passed to the forward43 method of [`Wav2Vec2Model`].44 hidden_size (`int`, *optional*, defaults to 768):45 Dimensionality of the encoder layers and the pooler layer.46 num_hidden_layers (`int`, *optional*, defaults to 12):47 Number of hidden layers in the Transformer encoder.48 num_attention_heads (`int`, *optional*, defaults to 12):49 Number of attention heads for each attention layer in the Transformer encoder.50 intermediate_size (`int`, *optional*, defaults to 3072):51 Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.52 hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):53 The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,54 `"relu"`, `"selu"` and `"gelu_new"` are supported.55 hidden_dropout (`float`, *optional*, defaults to 0.1):56 The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.57 activation_dropout (`float`, *optional*, defaults to 0.1):58 The dropout ratio for activations inside the fully connected layer.59 attention_dropout (`float`, *optional*, defaults to 0.1):60 The dropout ratio for the attention probabilities.61 final_dropout (`float`, *optional*, defaults to 0.1):62 The dropout probability for the final projection layer of [`Wav2Vec2ForCTC`].63 layerdrop (`float`, *optional*, defaults to 0.1):64 The LayerDrop probability. See the [LayerDrop paper](see https://huggingface.co/papers/1909.11556) for more65 details.66 initializer_range (`float`, *optional*, defaults to 0.02):67 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.68 layer_norm_eps (`float`, *optional*, defaults to 1e-12):69 The epsilon used by the layer normalization layers.70 feat_extract_norm (`str`, *optional*, defaults to `"group"`):71 The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group72 normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D73 convolutional layers.74 feat_proj_dropout (`float`, *optional*, defaults to 0.0):75 The dropout probability for output of the feature encoder.76 feat_extract_activation (`str, `optional`, defaults to `"gelu"`):77 The non-linear activation function (function or string) in the 1D convolutional layers of the feature78 extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.79 feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):80 The dropout probability for quantized feature encoder states.81 conv_dim (`tuple[int]` or `list[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):82 A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the83 feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.84 conv_stride (`tuple[int]` or `list[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):85 A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length86 of *conv_stride* defines the number of convolutional layers and has to match the length of *conv_dim*.87 conv_kernel (`tuple[int]` or `list[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):88 A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The89 length of *conv_kernel* defines the number of convolutional layers and has to match the length of90 *conv_dim*.91 conv_bias (`bool`, *optional*, defaults to `False`):92 Whether the 1D convolutional layers have a bias.93 num_conv_pos_embeddings (`int`, *optional*, defaults to 128):94 Number of convolutional positional embeddings. Defines the kernel size of 1D convolutional positional95 embeddings layer.96 num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16):97 Number of groups of 1D convolutional positional embeddings layer.98 do_stable_layer_norm (`bool`, *optional*, defaults to `False`):99 Whether to apply *stable* layer norm architecture of the Transformer encoder. `do_stable_layer_norm is100 True` corresponds to applying layer norm before the attention layer, whereas `do_stable_layer_norm is101 False` corresponds to applying layer norm after the attention layer.102 apply_spec_augment (`bool`, *optional*, defaults to `True`):103 Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see104 [SpecAugment: A Simple Data Augmentation Method for Automatic Speech105 Recognition](https://huggingface.co/papers/1904.08779).106 mask_time_prob (`float`, *optional*, defaults to 0.05):107 Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking108 procedure generates ''mask_time_prob*len(time_axis)/mask_time_length'' independent masks over the axis. If109 reasoning from the probability of each feature vector to be chosen as the start of the vector span to be110 masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the111 actual percentage of masked vectors. This is only relevant if `apply_spec_augment is True`.112 mask_time_length (`int`, *optional*, defaults to 10):113 Length of vector span along the time axis.114 mask_time_min_masks (`int`, *optional*, defaults to 2),:115 The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,116 irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <117 mask_time_min_masks''118 mask_feature_prob (`float`, *optional*, defaults to 0.0):119 Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The120 masking procedure generates ''mask_feature_prob*len(feature_axis)/mask_time_length'' independent masks over121 the axis. If reasoning from the probability of each feature vector to be chosen as the start of the vector122 span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap123 may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is124 True`.125 mask_feature_length (`int`, *optional*, defaults to 10):126 Length of vector span along the feature axis.127 mask_feature_min_masks (`int`, *optional*, defaults to 0),:128 The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time129 step, irrespectively of `mask_feature_prob`. Only relevant if130 ''mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks''131 num_codevectors_per_group (`int`, *optional*, defaults to 320):132 Number of entries in each quantization codebook (group).133 num_codevector_groups (`int`, *optional*, defaults to 2):134 Number of codevector groups for product codevector quantization.135 contrastive_logits_temperature (`float`, *optional*, defaults to 0.1):136 The temperature *kappa* in the contrastive loss.137 feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):138 The dropout probability for the output of the feature encoder that's used by the quantizer.139 num_negatives (`int`, *optional*, defaults to 100):140 Number of negative samples for the contrastive loss.141 codevector_dim (`int`, *optional*, defaults to 256):142 Dimensionality of the quantized feature vectors.143 proj_codevector_dim (`int`, *optional*, defaults to 256):144 Dimensionality of the final projection of both the quantized and the transformer features.145 diversity_loss_weight (`int`, *optional*, defaults to 0.1):146 The weight of the codebook diversity loss component.147 ctc_loss_reduction (`str`, *optional*, defaults to `"sum"`):148 Specifies the reduction to apply to the output of `torch.nn.CTCLoss`. Only relevant when training an149 instance of [`Wav2Vec2ForCTC`].150 ctc_zero_infinity (`bool`, *optional*, defaults to `False`):151 Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`. Infinite losses mainly152 occur when the inputs are too short to be aligned to the targets. Only relevant when training an instance153 of [`Wav2Vec2ForCTC`].154 use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):155 Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an156 instance of [`Wav2Vec2ForSequenceClassification`].157 classifier_proj_size (`int`, *optional*, defaults to 256):158 Dimensionality of the projection before token mean-pooling for classification.159 tdnn_dim (`tuple[int]` or `list[int]`, *optional*, defaults to `(512, 512, 512, 512, 1500)`):160 A tuple of integers defining the number of output channels of each 1D convolutional layer in the *TDNN*161 module of the *XVector* model. The length of *tdnn_dim* defines the number of *TDNN* layers.162 tdnn_kernel (`tuple[int]` or `list[int]`, *optional*, defaults to `(5, 3, 3, 1, 1)`):163 A tuple of integers defining the kernel size of each 1D convolutional layer in the *TDNN* module of the164 *XVector* model. The length of *tdnn_kernel* has to match the length of *tdnn_dim*.165 tdnn_dilation (`tuple[int]` or `list[int]`, *optional*, defaults to `(1, 2, 3, 1, 1)`):166 A tuple of integers defining the dilation factor of each 1D convolutional layer in *TDNN* module of the167 *XVector* model. The length of *tdnn_dilation* has to match the length of *tdnn_dim*.168 xvector_output_dim (`int`, *optional*, defaults to 512):169 Dimensionality of the *XVector* embedding vectors.170 add_adapter (`bool`, *optional*, defaults to `False`):171 Whether a convolutional network should be stacked on top of the Wav2Vec2 Encoder. Can be very useful for172 warm-starting Wav2Vec2 for SpeechEncoderDecoder models.173 adapter_kernel_size (`int`, *optional*, defaults to 3):174 Kernel size of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.175 adapter_stride (`int`, *optional*, defaults to 2):176 Stride of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.177 num_adapter_layers (`int`, *optional*, defaults to 3):178 Number of convolutional layers that should be used in the adapter network. Only relevant if `add_adapter is179 True`.180 adapter_attn_dim (`int`, *optional*):181 Dimension of the attention adapter weights to be used in each attention block. An example of a model using182 attention adapters is [facebook/mms-1b-all](https://huggingface.co/facebook/mms-1b-all).183 output_hidden_size (`int`, *optional*):184 Dimensionality of the encoder output layer. If not defined, this defaults to *hidden-size*. Only relevant185 if `add_adapter is True`.186 187 Example:188 189 ```python190 >>> from transformers import Wav2Vec2Config, Wav2Vec2Model191 192 >>> # Initializing a Wav2Vec2 facebook/wav2vec2-base-960h style configuration193 >>> configuration = Wav2Vec2Config()194 195 >>> # Initializing a model (with random weights) from the facebook/wav2vec2-base-960h style configuration196 >>> model = Wav2Vec2Model(configuration)197 198 >>> # Accessing the model configuration199 >>> configuration = model.config200 ```"""201 202 model_type = "wav2vec2"203 204 def __init__(205 self,206 vocab_size=32,207 hidden_size=768,208 num_hidden_layers=12,209 num_attention_heads=12,210 intermediate_size=3072,211 hidden_act="gelu",212 hidden_dropout=0.1,213 activation_dropout=0.1,214 attention_dropout=0.1,215 feat_proj_dropout=0.0,216 feat_quantizer_dropout=0.0,217 final_dropout=0.1,218 layerdrop=0.1,219 initializer_range=0.02,220 layer_norm_eps=1e-5,221 feat_extract_norm="group",222 feat_extract_activation="gelu",223 conv_dim=(512, 512, 512, 512, 512, 512, 512),224 conv_stride=(5, 2, 2, 2, 2, 2, 2),225 conv_kernel=(10, 3, 3, 3, 3, 2, 2),226 conv_bias=False,227 num_conv_pos_embeddings=128,228 num_conv_pos_embedding_groups=16,229 do_stable_layer_norm=False,230 apply_spec_augment=True,231 mask_time_prob=0.05,232 mask_time_length=10,233 mask_time_min_masks=2,234 mask_feature_prob=0.0,235 mask_feature_length=10,236 mask_feature_min_masks=0,237 num_codevectors_per_group=320,238 num_codevector_groups=2,239 contrastive_logits_temperature=0.1,240 num_negatives=100,241 codevector_dim=256,242 proj_codevector_dim=256,243 diversity_loss_weight=0.1,244 ctc_loss_reduction="sum",245 ctc_zero_infinity=False,246 use_weighted_layer_sum=False,247 classifier_proj_size=256,248 tdnn_dim=(512, 512, 512, 512, 1500),249 tdnn_kernel=(5, 3, 3, 1, 1),250 tdnn_dilation=(1, 2, 3, 1, 1),251 xvector_output_dim=512,252 pad_token_id=0,253 bos_token_id=1,254 eos_token_id=2,255 add_adapter=False,256 adapter_kernel_size=3,257 adapter_stride=2,258 num_adapter_layers=3,259 output_hidden_size=None,260 adapter_attn_dim=None,261 **kwargs,262 ):263 super().__init__(**kwargs, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id)264 self.hidden_size = hidden_size265 self.feat_extract_norm = feat_extract_norm266 self.feat_extract_activation = feat_extract_activation267 self.conv_dim = list(conv_dim)268 self.conv_stride = list(conv_stride)269 self.conv_kernel = list(conv_kernel)270 self.conv_bias = conv_bias271 self.num_conv_pos_embeddings = num_conv_pos_embeddings272 self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups273 self.num_feat_extract_layers = len(self.conv_dim)274 self.num_hidden_layers = num_hidden_layers275 self.intermediate_size = intermediate_size276 self.hidden_act = hidden_act277 self.num_attention_heads = num_attention_heads278 self.hidden_dropout = hidden_dropout279 self.attention_dropout = attention_dropout280 self.activation_dropout = activation_dropout281 self.feat_proj_dropout = feat_proj_dropout282 self.final_dropout = final_dropout283 self.layerdrop = layerdrop284 self.layer_norm_eps = layer_norm_eps285 self.initializer_range = initializer_range286 self.vocab_size = vocab_size287 self.do_stable_layer_norm = do_stable_layer_norm288 self.use_weighted_layer_sum = use_weighted_layer_sum289 290 if (291 (len(self.conv_stride) != self.num_feat_extract_layers)292 or (len(self.conv_kernel) != self.num_feat_extract_layers)293 or (len(self.conv_dim) != self.num_feat_extract_layers)294 ):295 raise ValueError(296 "Configuration for convolutional layers is incorrect. It is required that `len(config.conv_dim)` =="297 " `len(config.conv_stride)` == `len(config.conv_kernel)`, but is `len(config.conv_dim) ="298 f" {len(self.conv_dim)}`, `len(config.conv_stride) = {len(self.conv_stride)}`,"299 f" `len(config.conv_kernel) = {len(self.conv_kernel)}`."300 )301 302 # fine-tuning config parameters for SpecAugment: https://huggingface.co/papers/1904.08779303 self.apply_spec_augment = apply_spec_augment304 self.mask_time_prob = mask_time_prob305 self.mask_time_length = mask_time_length306 self.mask_time_min_masks = mask_time_min_masks307 self.mask_feature_prob = mask_feature_prob308 self.mask_feature_length = mask_feature_length309 self.mask_feature_min_masks = mask_feature_min_masks310 311 # parameters for pretraining with codevector quantized representations312 self.num_codevectors_per_group = num_codevectors_per_group313 self.num_codevector_groups = num_codevector_groups314 self.contrastive_logits_temperature = contrastive_logits_temperature315 self.feat_quantizer_dropout = feat_quantizer_dropout316 self.num_negatives = num_negatives317 self.codevector_dim = codevector_dim318 self.proj_codevector_dim = proj_codevector_dim319 self.diversity_loss_weight = diversity_loss_weight320 321 # ctc loss322 self.ctc_loss_reduction = ctc_loss_reduction323 self.ctc_zero_infinity = ctc_zero_infinity324 325 # adapter326 self.add_adapter = add_adapter327 self.adapter_kernel_size = adapter_kernel_size328 self.adapter_stride = adapter_stride329 self.num_adapter_layers = num_adapter_layers330 self.output_hidden_size = output_hidden_size or hidden_size331 self.adapter_attn_dim = adapter_attn_dim332 333 # SequenceClassification-specific parameter. Feel free to ignore for other classes.334 self.classifier_proj_size = classifier_proj_size335 336 # XVector-specific parameters. Feel free to ignore for other classes.337 self.tdnn_dim = list(tdnn_dim)338 self.tdnn_kernel = list(tdnn_kernel)339 self.tdnn_dilation = list(tdnn_dilation)340 self.xvector_output_dim = xvector_output_dim341 342 @property343 def inputs_to_logits_ratio(self):344 return functools.reduce(operator.mul, self.conv_stride, 1)345 346 347__all__ = ["Wav2Vec2Config"]348