Aluode/PerceptionLabPortable

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"""Data2VecText configuration"""16 17import math18 19from ...configuration_utils import PretrainedConfig20from ...utils import logging21 22 23logger = logging.get_logger(__name__)24 25 26class Data2VecAudioConfig(PretrainedConfig):27    r"""28    This is the configuration class to store the configuration of a [`Data2VecAudioModel`]. It is used to instantiate29    an Data2VecAudio 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 Data2VecAudio31    [facebook/data2vec-audio-base-960h](https://huggingface.co/facebook/data2vec-audio-base-960h) 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 37    Args:38        vocab_size (`int`, *optional*, defaults to 32):39            Vocabulary size of the Data2VecAudio model. Defines the number of different tokens that can be represented40            by the `inputs_ids` passed when calling [`Data2VecAudioModel`] or [`TFData2VecAudioModel`]. Vocabulary size41            of the model. Defines the different tokens that can be represented by the *inputs_ids* passed to the42            forward method of [`Data2VecAudioModel`].43        hidden_size (`int`, *optional*, defaults to 768):44            Dimensionality of the encoder layers and the pooler layer.45        num_hidden_layers (`int`, *optional*, defaults to 12):46            Number of hidden layers in the Transformer encoder.47        num_attention_heads (`int`, *optional*, defaults to 12):48            Number of attention heads for each attention layer in the Transformer encoder.49        intermediate_size (`int`, *optional*, defaults to 3072):50            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.51        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):52            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,53            `"relu"`, `"selu"` and `"gelu_new"` are supported.54        hidden_dropout (`float`, *optional*, defaults to 0.1):55            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.56        activation_dropout (`float`, *optional*, defaults to 0.1):57            The dropout ratio for activations inside the fully connected layer.58        attention_dropout (`float`, *optional*, defaults to 0.1):59            The dropout ratio for the attention probabilities.60        final_dropout (`float`, *optional*, defaults to 0.1):61            The dropout probability for the final projection layer of [`Data2VecAudioForCTC`].62        layerdrop (`float`, *optional*, defaults to 0.1):63            The LayerDrop probability. See the [LayerDrop paper](see https://huggingface.co/papers/1909.11556) for more64            details.65        initializer_range (`float`, *optional*, defaults to 0.02):66            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.67        layer_norm_eps (`float`, *optional*, defaults to 1e-12):68            The epsilon used by the layer normalization layers.69        feat_proj_dropout (`float`, *optional*, defaults to 0.0):70            The dropout probability for output of the feature encoder.71        feat_extract_activation (`str, `optional`, defaults to `"gelu"`):72            The non-linear activation function (function or string) in the 1D convolutional layers of the feature73            extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.74        conv_dim (`tuple[int]` or `list[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):75            A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the76            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.77        conv_stride (`tuple[int]` or `list[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):78            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length79            of *conv_stride* defines the number of convolutional layers and has to match the length of *conv_dim*.80        conv_kernel (`tuple[int]` or `list[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):81            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The82            length of *conv_kernel* defines the number of convolutional layers and has to match the length of83            *conv_dim*.84        conv_bias (`bool`, *optional*, defaults to `False`):85            Whether the 1D convolutional layers have a bias.86        num_conv_pos_embeddings (`int`, *optional*, defaults to 128):87            Number of convolutional positional embeddings. Defines the kernel size of 1D convolutional positional88            embeddings layer.89        num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16):90            Number of groups of 1D convolutional positional embeddings layer.91        mask_time_prob (`float`, *optional*, defaults to 0.05):92            Percentage (between 0 and 1) of all feature vectors along the time axis which will be masked. The masking93            procedure generates ''mask_time_prob*len(time_axis)/mask_time_length'' independent masks over the axis. If94            reasoning from the probability of each feature vector to be chosen as the start of the vector span to be95            masked, *mask_time_prob* should be `prob_vector_start*mask_time_length`. Note that overlap may decrease the96        mask_time_length (`int`, *optional*, defaults to 10):97            Length of vector span along the time axis.98        mask_time_min_masks (`int`, *optional*, defaults to 2),:99            The minimum number of masks of length `mask_feature_length` generated along the time axis, each time step,100            irrespectively of `mask_feature_prob`. Only relevant if ''mask_time_prob*len(time_axis)/mask_time_length <101            mask_time_min_masks''102        mask_feature_prob (`float`, *optional*, defaults to 0.0):103            Percentage (between 0 and 1) of all feature vectors along the feature axis which will be masked. The104            masking procedure generates ''mask_feature_prob*len(feature_axis)/mask_time_length'' independent masks over105            the axis. If reasoning from the probability of each feature vector to be chosen as the start of the vector106            span to be masked, *mask_feature_prob* should be `prob_vector_start*mask_feature_length`. Note that overlap107            may decrease the actual percentage of masked vectors. This is only relevant if `apply_spec_augment is108            True`.109        mask_feature_length (`int`, *optional*, defaults to 10):110            Length of vector span along the feature axis.111        mask_feature_min_masks (`int`, *optional*, defaults to 0),:112            The minimum number of masks of length `mask_feature_length` generated along the feature axis, each time113            step, irrespectively of `mask_feature_prob`. Only relevant if114            ''mask_feature_prob*len(feature_axis)/mask_feature_length < mask_feature_min_masks''115        ctc_loss_reduction (`str`, *optional*, defaults to `"sum"`):116            Specifies the reduction to apply to the output of `torch.nn.CTCLoss`. Only relevant when training an117            instance of [`Data2VecAudioForCTC`].118        ctc_zero_infinity (`bool`, *optional*, defaults to `False`):119            Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`. Infinite losses mainly120            occur when the inputs are too short to be aligned to the targets. Only relevant when training an instance121            of [`Data2VecAudioForCTC`].122        use_weighted_layer_sum (`bool`, *optional*, defaults to `False`):123            Whether to use a weighted average of layer outputs with learned weights. Only relevant when using an124            instance of [`Data2VecAudioForSequenceClassification`].125        classifier_proj_size (`int`, *optional*, defaults to 256):126            Dimensionality of the projection before token mean-pooling for classification.127        tdnn_dim (`tuple[int]` or `list[int]`, *optional*, defaults to `(512, 512, 512, 512, 1500)`):128            A tuple of integers defining the number of output channels of each 1D convolutional layer in the *TDNN*129            module of the *XVector* model. The length of *tdnn_dim* defines the number of *TDNN* layers.130        tdnn_kernel (`tuple[int]` or `list[int]`, *optional*, defaults to `(5, 3, 3, 1, 1)`):131            A tuple of integers defining the kernel size of each 1D convolutional layer in the *TDNN* module of the132            *XVector* model. The length of *tdnn_kernel* has to match the length of *tdnn_dim*.133        tdnn_dilation (`tuple[int]` or `list[int]`, *optional*, defaults to `(1, 2, 3, 1, 1)`):134            A tuple of integers defining the dilation factor of each 1D convolutional layer in *TDNN* module of the135            *XVector* model. The length of *tdnn_dilation* has to match the length of *tdnn_dim*.136        xvector_output_dim (`int`, *optional*, defaults to 512):137            Dimensionality of the *XVector* embedding vectors.138        add_adapter (`bool`, *optional*, defaults to `False`):139            Whether a convolutional network should be stacked on top of the Data2VecAudio Encoder. Can be very useful140            for warm-starting Data2VecAudio for SpeechEncoderDecoder models.141        adapter_kernel_size (`int`, *optional*, defaults to 3):142            Kernel size of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.143        adapter_stride (`int`, *optional*, defaults to 2):144            Stride of the convolutional layers in the adapter network. Only relevant if `add_adapter is True`.145        num_adapter_layers (`int`, *optional*, defaults to 3):146            Number of convolutional layers that should be used in the adapter network. Only relevant if `add_adapter is147            True`.148        output_hidden_size (`int`, *optional*):149            Dimensionality of the encoder output layer. If not defined, this defaults to *hidden-size*. Only relevant150            if `add_adapter is True`.151 152    Example:153 154    ```python155    >>> from transformers import Data2VecAudioConfig, Data2VecAudioModel156 157    >>> # Initializing a Data2VecAudio facebook/data2vec-audio-base-960h style configuration158    >>> configuration = Data2VecAudioConfig()159 160    >>> # Initializing a model (with random weights) from the facebook/data2vec-audio-base-960h style configuration161    >>> model = Data2VecAudioModel(configuration)162 163    >>> # Accessing the model configuration164    >>> configuration = model.config165    ```"""166 167    model_type = "data2vec-audio"168 169    def __init__(170        self,171        vocab_size=32,172        hidden_size=768,173        num_hidden_layers=12,174        num_attention_heads=12,175        intermediate_size=3072,176        hidden_act="gelu",177        hidden_dropout=0.1,178        activation_dropout=0.1,179        attention_dropout=0.1,180        feat_proj_dropout=0.0,181        final_dropout=0.1,182        layerdrop=0.1,183        initializer_range=0.02,184        layer_norm_eps=1e-5,185        feat_extract_activation="gelu",186        conv_dim=(512, 512, 512, 512, 512, 512, 512),187        conv_stride=(5, 2, 2, 2, 2, 2, 2),188        conv_kernel=(10, 3, 3, 3, 3, 2, 2),189        conv_bias=False,190        num_conv_pos_embedding_groups=16,191        conv_pos_kernel_size=19,192        num_conv_pos_embeddings=5,193        mask_time_prob=0.05,194        mask_time_length=10,195        mask_time_min_masks=2,196        mask_feature_prob=0.0,197        mask_feature_length=10,198        mask_feature_min_masks=0,199        ctc_loss_reduction="sum",200        ctc_zero_infinity=False,201        use_weighted_layer_sum=False,202        classifier_proj_size=256,203        tdnn_dim=(512, 512, 512, 512, 1500),204        tdnn_kernel=(5, 3, 3, 1, 1),205        tdnn_dilation=(1, 2, 3, 1, 1),206        xvector_output_dim=512,207        pad_token_id=0,208        bos_token_id=1,209        eos_token_id=2,210        add_adapter=False,211        adapter_kernel_size=3,212        adapter_stride=2,213        num_adapter_layers=3,214        output_hidden_size=None,215        **kwargs,216    ):217        super().__init__(**kwargs, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id)218        self.hidden_size = hidden_size219        self.feat_extract_activation = feat_extract_activation220        self.conv_dim = list(conv_dim)221        self.conv_stride = list(conv_stride)222        self.conv_kernel = list(conv_kernel)223        self.conv_bias = conv_bias224        self.num_conv_pos_embeddings = num_conv_pos_embeddings225        self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups226        self.conv_pos_kernel_size = conv_pos_kernel_size227        self.num_feat_extract_layers = len(self.conv_dim)228        self.num_hidden_layers = num_hidden_layers229        self.intermediate_size = intermediate_size230        self.hidden_act = hidden_act231        self.num_attention_heads = num_attention_heads232        self.hidden_dropout = hidden_dropout233        self.attention_dropout = attention_dropout234        self.activation_dropout = activation_dropout235        self.feat_proj_dropout = feat_proj_dropout236        self.final_dropout = final_dropout237        self.layerdrop = layerdrop238        self.layer_norm_eps = layer_norm_eps239        self.initializer_range = initializer_range240        self.vocab_size = vocab_size241        self.use_weighted_layer_sum = use_weighted_layer_sum242 243        if (244            (len(self.conv_stride) != self.num_feat_extract_layers)245            or (len(self.conv_kernel) != self.num_feat_extract_layers)246            or (len(self.conv_dim) != self.num_feat_extract_layers)247        ):248            raise ValueError(249                "Configuration for convolutional layers is incorrect. It is required that `len(config.conv_dim)` =="250                " `len(config.conv_stride)` == `len(config.conv_kernel)`, but is `len(config.conv_dim) ="251                f" {len(self.conv_dim)}`, `len(config.conv_stride) = {len(self.conv_stride)}`,"252                f" `len(config.conv_kernel) = {len(self.conv_kernel)}`."253            )254 255        # fine-tuning config parameters for SpecAugment: https://huggingface.co/papers/1904.08779256        self.mask_time_prob = mask_time_prob257        self.mask_time_length = mask_time_length258        self.mask_time_min_masks = mask_time_min_masks259        self.mask_feature_prob = mask_feature_prob260        self.mask_feature_length = mask_feature_length261        self.mask_feature_min_masks = mask_feature_min_masks262 263        # ctc loss264        self.ctc_loss_reduction = ctc_loss_reduction265        self.ctc_zero_infinity = ctc_zero_infinity266 267        # adapter268        self.add_adapter = add_adapter269        self.adapter_kernel_size = adapter_kernel_size270        self.adapter_stride = adapter_stride271        self.num_adapter_layers = num_adapter_layers272        self.output_hidden_size = output_hidden_size or hidden_size273 274        # SequenceClassification-specific parameter. Feel free to ignore for other classes.275        self.classifier_proj_size = classifier_proj_size276 277        # XVector-specific parameters. Feel free to ignore for other classes.278        self.tdnn_dim = list(tdnn_dim)279        self.tdnn_kernel = list(tdnn_kernel)280        self.tdnn_dilation = list(tdnn_dilation)281        self.xvector_output_dim = xvector_output_dim282 283    @property284    def inputs_to_logits_ratio(self):285        return math.prod(self.conv_stride)286 287 288__all__ = ["Data2VecAudioConfig"]289