# coding=utf-8 # Copyright 2021 The Fairseq Authors and the HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ PyTorch UniSpeech model. """ import math from dataclasses import dataclass from typing import Optional, Tuple, Union import numpy as np import torch import torch.utils.checkpoint from torch import nn from torch.nn import CrossEntropyLoss from ...activations import ACT2FN from ...deepspeed import is_deepspeed_zero3_enabled from ...file_utils import ( ModelOutput, add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, replace_return_docstrings, ) from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput from ...modeling_utils import PreTrainedModel from ...utils import logging from .configuration_unispeech import UniSpeechConfig logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "UniSpeechConfig" _PROCESSOR_FOR_DOC = "Wav2Vec2Processor" _CHECKPOINT_FOR_DOC = "microsoft/unispeech-large-1500h-cv" _SEQ_CLASS_CHECKPOINT = "microsoft/unispeech-large-1500h-cv" _SEQ_CLASS_PROCESSOR_FOR_DOC = "Wav2Vec2FeatureExtractor" _HIDDEN_STATES_START_POSITION = 2 UNISPEECH_PRETRAINED_MODEL_ARCHIVE_LIST = [ "microsoft/unispeech-large-1500h-cv", "microsoft/unispeech-large-multi-lingual-1500h-cv", # See all UniSpeech models at https://huggingface.co/models?filter=unispeech ] @dataclass class UniSpeechBaseModelOutput(ModelOutput): """ Output type of :class:`~transformers.UniSpeechBaseModelOutput`, with potential hidden states and attentions. Args: last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. extract_features (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, conv_dim[-1])`): Sequence of extracted feature vectors of the last convolutional layer of the model. hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``): Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ last_hidden_state: torch.FloatTensor = None extract_features: torch.FloatTensor = None hidden_states: Optional[Tuple[torch.FloatTensor]] = None attentions: Optional[Tuple[torch.FloatTensor]] = None @dataclass class UniSpeechForPreTrainingOutput(ModelOutput): """ Output type of :class:`~transformers.UniSpeechForPreTrainingOutput`, with potential hidden states and attentions. Args: loss (`optional`, returned when model is in train mode, ``torch.FloatTensor`` of shape :obj:`(1,)`): Total loss as the sum of the contrastive loss (L_m) and the diversity loss (L_d) as stated in the `official paper `__ . (classification) loss. projected_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.proj_codevector_dim)`): Hidden-states of the model projected to `config.proj_codevector_dim` that can be used to predict the masked projected quantized states. projected_quantized_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.proj_codevector_dim)`): Quantized extracted feature vectors projected to `config.proj_codevector_dim` representing the positive target vectors for contrastive loss. hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``): Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[torch.FloatTensor] = None projected_states: torch.FloatTensor = None projected_quantized_states: torch.FloatTensor = None codevector_perplexity: torch.FloatTensor = None hidden_states: Optional[Tuple[torch.FloatTensor]] = None attentions: Optional[Tuple[torch.FloatTensor]] = None # Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices def _compute_mask_indices( shape: Tuple[int, int], mask_prob: float, mask_length: int, attention_mask: Optional[torch.LongTensor] = None, min_masks: int = 0, ) -> np.ndarray: """ Computes random mask spans for a given shape. Used to implement `SpecAugment: A Simple Data Augmentation Method for ASR `__. Note that this method is not optimized to run on TPU and should be run on CPU as part of the preprocessing during training. Args: shape: The shape for which to compute masks. This should be of a tuple of size 2 where the first element is the batch size and the second element is the length of the axis to span. mask_prob: The percentage of the whole axis (between 0 and 1) which will be masked. The number of independently generated mask spans of length `mask_length` is computed by `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the actual percentage will be smaller. mask_length: size of the mask min_masks: minimum number of masked spans attention_mask: A (right-padded) attention mask which independently shortens the feature axis of each batch dimension. """ batch_size, sequence_length = shape if mask_length < 1: raise ValueError("`mask_length` has to be bigger than 0.") if mask_length > sequence_length: raise ValueError( f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}" f" and `sequence_length`: {sequence_length}`" ) # epsilon is used for probabilistic rounding epsilon = np.random.rand(1).item() def compute_num_masked_span(input_length): """Given input length, compute how many spans should be masked""" num_masked_span = int(mask_prob * input_length / mask_length + epsilon) num_masked_span = max(num_masked_span, min_masks) # make sure num masked indices <= sequence_length if num_masked_span * mask_length > sequence_length: num_masked_span = sequence_length // mask_length return num_masked_span # compute number of masked spans in batch input_lengths = ( attention_mask.sum(-1).detach().tolist() if attention_mask is not None else [sequence_length for _ in range(batch_size)] ) # SpecAugment mask to fill spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=np.bool) spec_aug_mask_idxs = [] max_num_masked_span = compute_num_masked_span(sequence_length) if max_num_masked_span == 0: return spec_aug_mask for input_length in input_lengths: # compute num of masked spans for this input num_masked_span = compute_num_masked_span(input_length) # get random indices to mask spec_aug_mask_idx = np.random.choice( np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False ) # pick first sampled index that will serve as a dummy index to pad vector # to ensure same dimension for all batches due to probabilistic rounding # Picking first sample just pads those vectors twice. dummy_mask_idx = spec_aug_mask_idx[0] spec_aug_mask_idx = np.concatenate( [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx] ) spec_aug_mask_idxs.append(spec_aug_mask_idx) spec_aug_mask_idxs = np.array(spec_aug_mask_idxs) # expand masked indices to masked spans spec_aug_mask_idxs = np.broadcast_to( spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length) ) spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length) # add offset to the starting indexes so that that indexes now create a span offsets = np.arange(mask_length)[None, None, :] offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape( batch_size, max_num_masked_span * mask_length ) spec_aug_mask_idxs = spec_aug_mask_idxs + offsets # scatter indices to mask np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1) return spec_aug_mask # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2NoLayerNormConvLayer with Wav2Vec2->UniSpeech class UniSpeechNoLayerNormConvLayer(nn.Module): def __init__(self, config, layer_id=0): super().__init__() self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 self.out_conv_dim = config.conv_dim[layer_id] self.conv = nn.Conv1d( self.in_conv_dim, self.out_conv_dim, kernel_size=config.conv_kernel[layer_id], stride=config.conv_stride[layer_id], bias=config.conv_bias, ) self.activation = ACT2FN[config.feat_extract_activation] def forward(self, hidden_states): hidden_states = self.conv(hidden_states) hidden_states = self.activation(hidden_states) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2LayerNormConvLayer with Wav2Vec2->UniSpeech class UniSpeechLayerNormConvLayer(nn.Module): def __init__(self, config, layer_id=0): super().__init__() self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 self.out_conv_dim = config.conv_dim[layer_id] self.conv = nn.Conv1d( self.in_conv_dim, self.out_conv_dim, kernel_size=config.conv_kernel[layer_id], stride=config.conv_stride[layer_id], bias=config.conv_bias, ) self.layer_norm = nn.LayerNorm(self.out_conv_dim, elementwise_affine=True) self.activation = ACT2FN[config.feat_extract_activation] def forward(self, hidden_states): hidden_states = self.conv(hidden_states) hidden_states = hidden_states.transpose(-2, -1) hidden_states = self.layer_norm(hidden_states) hidden_states = hidden_states.transpose(-2, -1) hidden_states = self.activation(hidden_states) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2GroupNormConvLayer with Wav2Vec2->UniSpeech class UniSpeechGroupNormConvLayer(nn.Module): def __init__(self, config, layer_id=0): super().__init__() self.in_conv_dim = config.conv_dim[layer_id - 1] if layer_id > 0 else 1 self.out_conv_dim = config.conv_dim[layer_id] self.conv = nn.Conv1d( self.in_conv_dim, self.out_conv_dim, kernel_size=config.conv_kernel[layer_id], stride=config.conv_stride[layer_id], bias=config.conv_bias, ) self.activation = ACT2FN[config.feat_extract_activation] self.layer_norm = nn.GroupNorm(num_groups=self.out_conv_dim, num_channels=self.out_conv_dim, affine=True) def forward(self, hidden_states): hidden_states = self.conv(hidden_states) hidden_states = self.layer_norm(hidden_states) hidden_states = self.activation(hidden_states) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2PositionalConvEmbedding with Wav2Vec2->UniSpeech class UniSpeechPositionalConvEmbedding(nn.Module): def __init__(self, config): super().__init__() self.conv = nn.Conv1d( config.hidden_size, config.hidden_size, kernel_size=config.num_conv_pos_embeddings, padding=config.num_conv_pos_embeddings // 2, groups=config.num_conv_pos_embedding_groups, ) if is_deepspeed_zero3_enabled(): import deepspeed with deepspeed.zero.GatheredParameters(self.conv.weight, modifier_rank=0): self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) deepspeed.zero.register_external_parameter(self, self.conv.weight_v) deepspeed.zero.register_external_parameter(self, self.conv.weight_g) else: self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) self.padding = UniSpeechSamePadLayer(config.num_conv_pos_embeddings) self.activation = ACT2FN[config.feat_extract_activation] def forward(self, hidden_states): hidden_states = hidden_states.transpose(1, 2) hidden_states = self.conv(hidden_states) hidden_states = self.padding(hidden_states) hidden_states = self.activation(hidden_states) hidden_states = hidden_states.transpose(1, 2) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2SamePadLayer with Wav2Vec2->UniSpeech class UniSpeechSamePadLayer(nn.Module): def __init__(self, num_conv_pos_embeddings): super().__init__() self.num_pad_remove = 1 if num_conv_pos_embeddings % 2 == 0 else 0 def forward(self, hidden_states): if self.num_pad_remove > 0: hidden_states = hidden_states[:, :, : -self.num_pad_remove] return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->UniSpeech class UniSpeechFeatureExtractor(nn.Module): """Construct the features from raw audio waveform""" def __init__(self, config): super().__init__() if config.feat_extract_norm == "group": conv_layers = [UniSpeechGroupNormConvLayer(config, layer_id=0)] + [ UniSpeechNoLayerNormConvLayer(config, layer_id=i + 1) for i in range(config.num_feat_extract_layers - 1) ] elif config.feat_extract_norm == "layer": conv_layers = [ UniSpeechLayerNormConvLayer(config, layer_id=i) for i in range(config.num_feat_extract_layers) ] else: raise ValueError( f"`config.feat_extract_norm` is {config.feat_extract_norm}, but has to be one of ['group', 'layer']" ) self.conv_layers = nn.ModuleList(conv_layers) self.gradient_checkpointing = False self._requires_grad = True def _freeze_parameters(self): for param in self.parameters(): param.requires_grad = False self._requires_grad = False def forward(self, input_values): hidden_states = input_values[:, None] # make sure hidden_states require grad for gradient_checkpointing if self._requires_grad and self.training: hidden_states.requires_grad = True for conv_layer in self.conv_layers: if self._requires_grad and self.gradient_checkpointing and self.training: def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs) return custom_forward hidden_states = torch.utils.checkpoint.checkpoint( create_custom_forward(conv_layer), hidden_states, ) else: hidden_states = conv_layer(hidden_states) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureProjection with Wav2Vec2->UniSpeech class UniSpeechFeatureProjection(nn.Module): def __init__(self, config): super().__init__() self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps) self.projection = nn.Linear(config.conv_dim[-1], config.hidden_size) self.dropout = nn.Dropout(config.feat_proj_dropout) def forward(self, hidden_states): # non-projected hidden states are needed for quantization norm_hidden_states = self.layer_norm(hidden_states) hidden_states = self.projection(norm_hidden_states) hidden_states = self.dropout(hidden_states) return hidden_states, norm_hidden_states # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->UniSpeech class UniSpeechAttention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__( self, embed_dim: int, num_heads: int, dropout: float = 0.0, is_decoder: bool = False, bias: bool = True, ): super().__init__() self.embed_dim = embed_dim self.num_heads = num_heads self.dropout = dropout self.head_dim = embed_dim // num_heads if (self.head_dim * num_heads) != self.embed_dim: raise ValueError( f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" f" and `num_heads`: {num_heads})." ) self.scaling = self.head_dim ** -0.5 self.is_decoder = is_decoder self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() def forward( self, hidden_states: torch.Tensor, key_value_states: Optional[torch.Tensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, attention_mask: Optional[torch.Tensor] = None, layer_head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: """Input shape: Batch x Time x Channel""" # if key_value_states are provided this layer is used as a cross-attention layer # for the decoder is_cross_attention = key_value_states is not None bsz, tgt_len, _ = hidden_states.size() # get query proj query_states = self.q_proj(hidden_states) * self.scaling # get key, value proj if is_cross_attention and past_key_value is not None: # reuse k,v, cross_attentions key_states = past_key_value[0] value_states = past_key_value[1] elif is_cross_attention: # cross_attentions key_states = self._shape(self.k_proj(key_value_states), -1, bsz) value_states = self._shape(self.v_proj(key_value_states), -1, bsz) elif past_key_value is not None: # reuse k, v, self_attention key_states = self._shape(self.k_proj(hidden_states), -1, bsz) value_states = self._shape(self.v_proj(hidden_states), -1, bsz) key_states = torch.cat([past_key_value[0], key_states], dim=2) value_states = torch.cat([past_key_value[1], value_states], dim=2) else: # self_attention key_states = self._shape(self.k_proj(hidden_states), -1, bsz) value_states = self._shape(self.v_proj(hidden_states), -1, bsz) if self.is_decoder: # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. # Further calls to cross_attention layer can then reuse all cross-attention # key/value_states (first "if" case) # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of # all previous decoder key/value_states. Further calls to uni-directional self-attention # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) # if encoder bi-directional self-attention `past_key_value` is always `None` past_key_value = (key_states, value_states) proj_shape = (bsz * self.num_heads, -1, self.head_dim) query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) key_states = key_states.view(*proj_shape) value_states = value_states.view(*proj_shape) src_len = key_states.size(1) attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): raise ValueError( f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}" ) if attention_mask is not None: if attention_mask.size() != (bsz, 1, tgt_len, src_len): raise ValueError( f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" ) attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) attn_weights = nn.functional.softmax(attn_weights, dim=-1) if layer_head_mask is not None: if layer_head_mask.size() != (self.num_heads,): raise ValueError( f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}" ) attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) if output_attentions: # this operation is a bit awkward, but it's required to # make sure that attn_weights keeps its gradient. # In order to do so, attn_weights have to be reshaped # twice and have to be reused in the following attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) else: attn_weights_reshaped = None attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) attn_output = torch.bmm(attn_probs, value_states) if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): raise ValueError( f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}" ) attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) attn_output = attn_output.transpose(1, 2) # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be # partitioned aross GPUs when using tensor-parallelism. attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) attn_output = self.out_proj(attn_output) return attn_output, attn_weights_reshaped, past_key_value # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeedForward with Wav2Vec2->UniSpeech class UniSpeechFeedForward(nn.Module): def __init__(self, config): super().__init__() self.intermediate_dropout = nn.Dropout(config.activation_dropout) self.intermediate_dense = nn.Linear(config.hidden_size, config.intermediate_size) if isinstance(config.hidden_act, str): self.intermediate_act_fn = ACT2FN[config.hidden_act] else: self.intermediate_act_fn = config.hidden_act self.output_dense = nn.Linear(config.intermediate_size, config.hidden_size) self.output_dropout = nn.Dropout(config.hidden_dropout) def forward(self, hidden_states): hidden_states = self.intermediate_dense(hidden_states) hidden_states = self.intermediate_act_fn(hidden_states) hidden_states = self.intermediate_dropout(hidden_states) hidden_states = self.output_dense(hidden_states) hidden_states = self.output_dropout(hidden_states) return hidden_states # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayer with Wav2Vec2->UniSpeech class UniSpeechEncoderLayer(nn.Module): def __init__(self, config): super().__init__() self.attention = UniSpeechAttention( embed_dim=config.hidden_size, num_heads=config.num_attention_heads, dropout=config.attention_dropout, is_decoder=False, ) self.dropout = nn.Dropout(config.hidden_dropout) self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.feed_forward = UniSpeechFeedForward(config) self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) def forward(self, hidden_states, attention_mask=None, output_attentions=False): attn_residual = hidden_states hidden_states, attn_weights, _ = self.attention( hidden_states, attention_mask=attention_mask, output_attentions=output_attentions ) hidden_states = self.dropout(hidden_states) hidden_states = attn_residual + hidden_states hidden_states = self.layer_norm(hidden_states) hidden_states = hidden_states + self.feed_forward(hidden_states) hidden_states = self.final_layer_norm(hidden_states) outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) return outputs # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderLayerStableLayerNorm with Wav2Vec2->UniSpeech class UniSpeechEncoderLayerStableLayerNorm(nn.Module): def __init__(self, config): super().__init__() self.attention = UniSpeechAttention( embed_dim=config.hidden_size, num_heads=config.num_attention_heads, dropout=config.attention_dropout, is_decoder=False, ) self.dropout = nn.Dropout(config.hidden_dropout) self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.feed_forward = UniSpeechFeedForward(config) self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) def forward(self, hidden_states, attention_mask=None, output_attentions=False): attn_residual = hidden_states hidden_states = self.layer_norm(hidden_states) hidden_states, attn_weights, _ = self.attention( hidden_states, attention_mask=attention_mask, output_attentions=output_attentions ) hidden_states = self.dropout(hidden_states) hidden_states = attn_residual + hidden_states hidden_states = hidden_states + self.feed_forward(self.final_layer_norm(hidden_states)) outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) return outputs # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Encoder with Wav2Vec2->UniSpeech class UniSpeechEncoder(nn.Module): def __init__(self, config): super().__init__() self.config = config self.pos_conv_embed = UniSpeechPositionalConvEmbedding(config) self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.dropout = nn.Dropout(config.hidden_dropout) self.layers = nn.ModuleList([UniSpeechEncoderLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, hidden_states, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True, ): all_hidden_states = () if output_hidden_states else None all_self_attentions = () if output_attentions else None if attention_mask is not None: # make sure padded tokens output 0 hidden_states[~attention_mask] = 0.0 # extend attention_mask attention_mask = (1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)) * -10000.0 attention_mask = attention_mask.expand( attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1] ) position_embeddings = self.pos_conv_embed(hidden_states) hidden_states = hidden_states + position_embeddings hidden_states = self.layer_norm(hidden_states) hidden_states = self.dropout(hidden_states) deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled() for layer in self.layers: if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) dropout_probability = np.random.uniform(0, 1) skip_the_layer = True if self.training and (dropout_probability < self.config.layerdrop) else False if not skip_the_layer or deepspeed_zero3_is_enabled: # under deepspeed zero3 all gpus must run in sync if self.gradient_checkpointing and self.training: # create gradient checkpointing function def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs, output_attentions) return custom_forward layer_outputs = torch.utils.checkpoint.checkpoint( create_custom_forward(layer), hidden_states, attention_mask, ) else: layer_outputs = layer( hidden_states, attention_mask=attention_mask, output_attentions=output_attentions ) hidden_states = layer_outputs[0] if skip_the_layer: layer_outputs = (None, None) if output_attentions: all_self_attentions = all_self_attentions + (layer_outputs[1],) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_self_attentions, ) # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2EncoderStableLayerNorm with Wav2Vec2->UniSpeech class UniSpeechEncoderStableLayerNorm(nn.Module): def __init__(self, config): super().__init__() self.config = config self.pos_conv_embed = UniSpeechPositionalConvEmbedding(config) self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.dropout = nn.Dropout(config.hidden_dropout) self.layers = nn.ModuleList( [UniSpeechEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)] ) self.gradient_checkpointing = False def forward( self, hidden_states, attention_mask=None, output_attentions=False, output_hidden_states=False, return_dict=True, ): all_hidden_states = () if output_hidden_states else None all_self_attentions = () if output_attentions else None if attention_mask is not None: # make sure padded tokens are not attended to hidden_states[~attention_mask] = 0 # extend attention_mask attention_mask = (1.0 - attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)) * -10000.0 attention_mask = attention_mask.expand( attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1] ) position_embeddings = self.pos_conv_embed(hidden_states) hidden_states = hidden_states + position_embeddings hidden_states = self.dropout(hidden_states) deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled() for layer in self.layers: if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) dropout_probability = np.random.uniform(0, 1) skip_the_layer = True if self.training and (dropout_probability < self.config.layerdrop) else False if not skip_the_layer or deepspeed_zero3_is_enabled: # under deepspeed zero3 all gpus must run in sync # XXX: could optimize this like synced_gpus in generate_utils but not sure if it's worth the code complication if self.gradient_checkpointing and self.training: # create gradient checkpointing function def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs, output_attentions) return custom_forward layer_outputs = torch.utils.checkpoint.checkpoint( create_custom_forward(layer), hidden_states, attention_mask, ) else: layer_outputs = layer( hidden_states, attention_mask=attention_mask, output_attentions=output_attentions ) hidden_states = layer_outputs[0] if skip_the_layer: layer_outputs = (None, None) if output_attentions: all_self_attentions = all_self_attentions + (layer_outputs[1],) hidden_states = self.layer_norm(hidden_states) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_self_attentions, ) class UniSpeechGumbelVectorQuantizer(nn.Module): """ Vector quantization using gumbel softmax. See `CATEGORICAL REPARAMETERIZATION WITH GUMBEL-SOFTMAX `__ for more information. """ def __init__(self, config): super().__init__() self.num_groups = config.num_codevector_groups self.num_vars = config.num_codevectors_per_group assert ( config.codevector_dim % self.num_groups == 0 ), f"`config.codevector_dim {config.codevector_dim} must be divisible by `config.num_codevector_groups` {self.num_groups} for concatenation" # storage for codebook variables (codewords) self.codevectors = nn.Parameter( torch.FloatTensor(1, self.num_groups * self.num_vars, config.codevector_dim // self.num_groups) ) self.weight_proj = nn.Linear(config.conv_dim[-1], self.num_groups * self.num_vars) # can be decayed for training self.temperature = 2 @staticmethod def _compute_perplexity(probs): marginal_probs = probs.mean(dim=0) perplexity = torch.exp(-torch.sum(marginal_probs * torch.log(marginal_probs + 1e-7), dim=-1)).sum() return perplexity def forward(self, hidden_states): batch_size, sequence_length, hidden_size = hidden_states.shape # project to codevector dim hidden_states = self.weight_proj(hidden_states) hidden_states = hidden_states.view(batch_size * sequence_length * self.num_groups, -1) if self.training: # sample code vector probs via gumbel in differentiateable way codevector_probs = nn.functional.gumbel_softmax( hidden_states.float(), tau=self.temperature, hard=True ).type_as(hidden_states) # compute perplexity codevector_soft_dist = torch.softmax( hidden_states.view(batch_size * sequence_length, self.num_groups, -1).float(), dim=-1 ) perplexity = self._compute_perplexity(codevector_soft_dist) else: # take argmax in non-differentiable way # comptute hard codevector distribution (one hot) codevector_idx = hidden_states.argmax(dim=-1) codevector_probs = hidden_states.new_zeros(*hidden_states.shape).scatter_( -1, codevector_idx.view(-1, 1), 1.0 ) codevector_probs = codevector_probs.view(batch_size * sequence_length, self.num_groups, -1) perplexity = self._compute_perplexity(codevector_probs) codevector_probs = codevector_probs.view(batch_size * sequence_length, -1) # use probs to retrieve codevectors codevectors_per_group = codevector_probs.unsqueeze(-1) * self.codevectors codevectors = codevectors_per_group.view(batch_size * sequence_length, self.num_groups, self.num_vars, -1) codevectors = codevectors.sum(-2).view(batch_size, sequence_length, -1) return codevectors, perplexity class UniSpeechPreTrainedModel(PreTrainedModel): """ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. """ config_class = UniSpeechConfig base_model_prefix = "unispeech" _keys_to_ignore_on_load_missing = [r"position_ids"] supports_gradient_checkpointing = True def _init_weights(self, module): """Initialize the weights""" # gumbel softmax requires special init if isinstance(module, UniSpeechGumbelVectorQuantizer): module.weight_proj.weight.data.normal_(mean=0.0, std=1) module.weight_proj.bias.data.zero_() nn.init.uniform_(module.codevectors) elif isinstance(module, UniSpeechPositionalConvEmbedding): nn.init.normal_( module.conv.weight, mean=0, std=2 * math.sqrt(1 / (module.conv.kernel_size[0] * module.conv.in_channels)), ) nn.init.constant_(module.conv.bias, 0) elif isinstance(module, UniSpeechFeatureProjection): k = math.sqrt(1 / module.projection.in_features) nn.init.uniform_(module.projection.weight, a=-k, b=k) nn.init.uniform_(module.projection.bias, a=-k, b=k) elif isinstance(module, nn.Linear): module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)): module.bias.data.zero_() module.weight.data.fill_(1.0) elif isinstance(module, nn.Conv1d): nn.init.kaiming_normal_(module.weight) if module.bias is not None: k = math.sqrt(module.groups / (module.in_channels * module.kernel_size[0])) nn.init.uniform_(module.bias, a=-k, b=k) def _get_feat_extract_output_lengths(self, input_lengths: Union[torch.LongTensor, int]): """ Computes the output length of the convolutional layers """ def _conv_out_length(input_length, kernel_size, stride): # 1D convolutional layer output length formula taken # from https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html return (input_length - kernel_size) // stride + 1 for kernel_size, stride in zip(self.config.conv_kernel, self.config.conv_stride): input_lengths = _conv_out_length(input_lengths, kernel_size, stride) return input_lengths def _get_feature_vector_attention_mask(self, feature_vector_length: int, attention_mask: torch.LongTensor): # Effectively attention_mask.sum(-1), but not inplace to be able to run # on inference mode. non_padded_lengths = attention_mask.cumsum(dim=-1)[:, -1] output_lengths = self._get_feat_extract_output_lengths(non_padded_lengths).to(torch.long) batch_size = attention_mask.shape[0] attention_mask = torch.zeros( (batch_size, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device ) # these two operations makes sure that all values before the output lengths idxs are attended to attention_mask[(torch.arange(attention_mask.shape[0], device=attention_mask.device), output_lengths - 1)] = 1 attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool() return attention_mask def _set_gradient_checkpointing(self, module, value=False): if isinstance(module, (UniSpeechEncoder, UniSpeechEncoderStableLayerNorm, UniSpeechFeatureExtractor)): module.gradient_checkpointing = value UNISPEECH_START_DOCSTRING = r""" UniSpeech was proposed in `UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data `__ by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang. This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving etc.). This model is a PyTorch `torch.nn.Module `_ sub-class. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config (:class:`~transformers.UniSpeechConfig`): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights. """ UNISPEECH_INPUTS_DOCSTRING = r""" Args: input_values (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`): Float values of input raw speech waveform. Values can be obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into `input_values`, the :class:`~transformers.UniSpeechProcessor` should be used for padding and conversion into a tensor of type `torch.FloatTensor`. See :meth:`transformers.UniSpeechProcessor.__call__` for details. attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): Mask to avoid performing convolution and attention on padding token indices. Mask values selected in ``[0, 1]``: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. `What are attention masks? <../glossary.html#attention-mask>`__ .. warning:: :obj:`attention_mask` should only be passed if the corresponding processor has ``config.return_attention_mask == True``. For all models whose processor has ``config.return_attention_mask == False``, :obj:`attention_mask` should **not** be passed to avoid degraded performance when doing batched inference. For such models :obj:`input_values` should simply be padded with 0 and passed without :obj:`attention_mask`. Be aware that these models also yield slightly different results depending on whether :obj:`input_values` is padded or not. output_attentions (:obj:`bool`, `optional`): Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned tensors for more detail. output_hidden_states (:obj:`bool`, `optional`): Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for more detail. return_dict (:obj:`bool`, `optional`): Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple. """ @add_start_docstrings( "The bare UniSpeech Model transformer outputting raw hidden-states without any specific head on top.", UNISPEECH_START_DOCSTRING, ) class UniSpeechModel(UniSpeechPreTrainedModel): def __init__(self, config: UniSpeechConfig): super().__init__(config) self.config = config self.feature_extractor = UniSpeechFeatureExtractor(config) self.feature_projection = UniSpeechFeatureProjection(config) self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_()) if config.do_stable_layer_norm: self.encoder = UniSpeechEncoderStableLayerNorm(config) else: self.encoder = UniSpeechEncoder(config) # Initialize weights and apply final processing self.post_init() # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Model._mask_hidden_states def _mask_hidden_states( self, hidden_states: torch.FloatTensor, mask_time_indices: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.LongTensor] = None, ): """ Masks extracted features along time axis and/or along feature axis according to `SpecAugment `__ . """ # `config.apply_spec_augment` can set masking to False if not getattr(self.config, "apply_spec_augment", True): return hidden_states # generate indices & apply SpecAugment along time axis batch_size, sequence_length, hidden_size = hidden_states.size() if mask_time_indices is not None: # apply SpecAugment along time axis with given mask_time_indices hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype) elif self.config.mask_time_prob > 0 and self.training: mask_time_indices = _compute_mask_indices( (batch_size, sequence_length), mask_prob=self.config.mask_time_prob, mask_length=self.config.mask_time_length, attention_mask=attention_mask, min_masks=self.config.mask_time_min_masks, ) mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool) hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype) if self.config.mask_feature_prob > 0 and self.training: # generate indices & apply SpecAugment along feature axis mask_feature_indices = _compute_mask_indices( (batch_size, hidden_size), mask_prob=self.config.mask_feature_prob, mask_length=self.config.mask_feature_length, min_masks=self.config.mask_feature_min_masks, ) mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool) mask_feature_indices = mask_feature_indices[:, None].expand(-1, sequence_length, -1) hidden_states[mask_feature_indices] = 0 return hidden_states @add_start_docstrings_to_model_forward(UNISPEECH_INPUTS_DOCSTRING) @add_code_sample_docstrings( processor_class=_PROCESSOR_FOR_DOC, checkpoint=_CHECKPOINT_FOR_DOC, output_type=UniSpeechBaseModelOutput, config_class=_CONFIG_FOR_DOC, modality="audio", ) def forward( self, input_values, attention_mask=None, mask_time_indices=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict extract_features = self.feature_extractor(input_values) extract_features = extract_features.transpose(1, 2) if attention_mask is not None: # compute reduced attention_mask corresponding to feature vectors attention_mask = self._get_feature_vector_attention_mask(extract_features.shape[1], attention_mask) hidden_states, extract_features = self.feature_projection(extract_features) hidden_states = self._mask_hidden_states( hidden_states, mask_time_indices=mask_time_indices, attention_mask=attention_mask ) encoder_outputs = self.encoder( hidden_states, attention_mask=attention_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = encoder_outputs[0] if not return_dict: return (hidden_states, extract_features) + encoder_outputs[1:] return UniSpeechBaseModelOutput( last_hidden_state=hidden_states, extract_features=extract_features, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) @add_start_docstrings( """UniSpeech Model with a vector-quantization module and ctc loss for pre-training. """, UNISPEECH_START_DOCSTRING ) class UniSpeechForPreTraining(UniSpeechPreTrainedModel): def __init__(self, config: UniSpeechConfig): super().__init__(config) self.unispeech = UniSpeechModel(config) self.dropout_features = nn.Dropout(config.feat_quantizer_dropout) self.quantizer = UniSpeechGumbelVectorQuantizer(config) self.project_q = nn.Linear(config.codevector_dim, config.proj_codevector_dim) self.project_hid = nn.Linear(config.proj_codevector_dim, config.hidden_size) self.ctc_proj = nn.Linear(config.hidden_size, config.num_ctc_classes) self.dropout = nn.Dropout(config.final_dropout) # Initialize weights and apply final processing self.post_init() def set_gumbel_temperature(self, temperature: int): """ Set the Gumbel softmax temperature to a given value. Only necessary for training """ self.quantizer.temperature = temperature def freeze_feature_extractor(self): """ Calling this function will disable the gradient computation for the feature extractor so that its parameters will not be updated during training. """ self.unispeech.feature_extractor._freeze_parameters() @staticmethod def compute_contrastive_logits( target_features: torch.FloatTensor, negative_features: torch.FloatTensor, predicted_features: torch.FloatTensor, temperature: int = 1, ): """ Compute logits for contrastive loss based using cosine similarity as the distance measure between :obj:`[positive_feature, negative_features]` and :obj:`[predicted_features]`. Additionally, temperature can be applied. """ target_features = torch.cat([target_features, negative_features], dim=0) logits = torch.cosine_similarity(predicted_features.float(), target_features.float(), dim=-1) logits = logits.type_as(target_features) # apply temperature logits = logits / temperature return logits @add_start_docstrings_to_model_forward(UNISPEECH_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=UniSpeechForPreTrainingOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_values, attention_mask=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" mask_time_indices (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): Indices to mask extracted features for contrastive loss. When in training mode, model learns to predict masked extracted features in `config.proj_codevector_dim` space. sampled_negative_indices (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, sequence_length, num_negatives)`, `optional`): Indices indicating which quantized target vectors are used as negative sampled vectors in contrastive loss. Required input for pre-training. Returns: Example:: >>> import torch >>> from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForPreTraining >>> from transformers.models.wav2vec2.modeling_wav2vec2 import _compute_mask_indices >>> from datasets import load_dataset >>> import soundfile as sf >>> feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("patrickvonplaten/wav2vec2-base") >>> model = Wav2Vec2ForPreTraining.from_pretrained("patrickvonplaten/wav2vec2-base") >>> def map_to_array(batch): ... speech, _ = sf.read(batch["file"]) ... batch["speech"] = speech ... return batch >>> ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation") >>> ds = ds.map(map_to_array) >>> input_values = feature_extractor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1 >>> # compute masked indices >>> batch_size, raw_sequence_length = input_values.shape >>> sequence_length = model._get_feat_extract_output_lengths(raw_sequence_length) >>> mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob=0.2, mask_length=2, device=model.device) >>> with torch.no_grad(): ... outputs = model(input_values, mask_time_indices=mask_time_indices) >>> # compute cosine similarity between predicted (=projected_states) and target (=projected_quantized_states) >>> cosine_sim = torch.cosine_similarity( ... outputs.projected_states, outputs.projected_quantized_states, dim=-1 ... ) >>> # show that cosine similarity is much higher than random >>> assert cosine_sim[mask_time_indices].mean() > 0.5 >>> # for contrastive loss training model should be put into train mode >>> model.train() >>> loss = model(input_values, mask_time_indices=mask_time_indices).loss """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.unispeech( input_values, attention_mask=attention_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) transformer_features = outputs[0] # quantize all (unmasked) extracted features and project to final vq dim extract_features = self.dropout_features(outputs[1]) quantized_features, codevector_perplexity = self.quantizer(extract_features) # project quantized features twice quantized_features = self.project_q(quantized_features) quantized_features = self.project_hid(quantized_features) prob_replace_matrix = torch.empty(transformer_features.size(0), transformer_features.size(1)).fill_( self.config.replace_prob ) prob_replace_matrix = prob_replace_matrix.transpose(0, 1) sampled_replace_matrix = torch.bernoulli(prob_replace_matrix).bool().to(transformer_features.device) sampled_replace_matrix = sampled_replace_matrix.transpose(0, 1) sampled_replace_matrix = sampled_replace_matrix.unsqueeze(-1) logits = transformer_features.masked_fill(sampled_replace_matrix, 0.0) + ( quantized_features.masked_fill(~sampled_replace_matrix, 0.0) ) # project to ctc units logits = self.dropout(logits) logits = self.ctc_proj(logits) # TODO(PVP) - add negative sampling & loss computation loss = None if not return_dict: if loss is not None: return (loss, transformer_features, quantized_features, codevector_perplexity) + outputs[2:] return (transformer_features, quantized_features, codevector_perplexity) + outputs[2:] return UniSpeechForPreTrainingOutput( loss=loss, projected_states=transformer_features, projected_quantized_states=quantized_features, codevector_perplexity=codevector_perplexity, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) @add_start_docstrings( """UniSpeech Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC). """, UNISPEECH_START_DOCSTRING, ) # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2ForCTC with Wav2Vec2->UniSpeech, wav2vec2->unispeech, WAV_2_VEC_2->UNISPEECH class UniSpeechForCTC(UniSpeechPreTrainedModel): def __init__(self, config): super().__init__(config) self.unispeech = UniSpeechModel(config) self.dropout = nn.Dropout(config.final_dropout) if config.vocab_size is None: raise ValueError( f"You are trying to instantiate {self.__class__} with a configuration that " "does not define the vocabulary size of the language model head. Please " "instantiate the model as follows: `UniSpeechForCTC.from_pretrained(..., vocab_size=vocab_size)`. " "or define `vocab_size` of your model's configuration." ) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size) # Initialize weights and apply final processing self.post_init() def freeze_feature_extractor(self): """ Calling this function will disable the gradient computation for the feature extractor so that its parameter will not be updated during training. """ self.unispeech.feature_extractor._freeze_parameters() @add_start_docstrings_to_model_forward(UNISPEECH_INPUTS_DOCSTRING) @add_code_sample_docstrings( processor_class=_PROCESSOR_FOR_DOC, checkpoint=_CHECKPOINT_FOR_DOC, output_type=CausalLMOutput, config_class=_CONFIG_FOR_DOC, ) def forward( self, input_values, attention_mask=None, output_attentions=None, output_hidden_states=None, return_dict=None, labels=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_length)`, `optional`): Labels for connectionist temporal classification. Note that ``target_length`` has to be smaller or equal to the sequence length of the output logits. Indices are selected in ``[-100, 0, ..., config.vocab_size - 1]``. All labels set to ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size - 1]``. """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.unispeech( input_values, attention_mask=attention_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = outputs[0] hidden_states = self.dropout(hidden_states) logits = self.lm_head(hidden_states) loss = None if labels is not None: if labels.max() >= self.config.vocab_size: raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}") # retrieve loss input_lengths from attention_mask attention_mask = ( attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long) ) input_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long) # assuming that padded tokens are filled with -100 # when not being attended to labels_mask = labels >= 0 target_lengths = labels_mask.sum(-1) flattened_targets = labels.masked_select(labels_mask) # ctc_loss doesn't support fp16 log_probs = nn.functional.log_softmax(logits, dim=-1, dtype=torch.float32).transpose(0, 1) with torch.backends.cudnn.flags(enabled=False): loss = nn.functional.ctc_loss( log_probs, flattened_targets, input_lengths, target_lengths, blank=self.config.pad_token_id, reduction=self.config.ctc_loss_reduction, zero_infinity=self.config.ctc_zero_infinity, ) if not return_dict: output = (logits,) + outputs[_HIDDEN_STATES_START_POSITION:] return ((loss,) + output) if loss is not None else output return CausalLMOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions ) @add_start_docstrings( """ UniSpeech Model with a sequence classification head on top (a linear layer over the pooled output) for tasks like SUPERB Keyword Spotting. """, UNISPEECH_START_DOCSTRING, ) # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2ForSequenceClassification with Wav2Vec2->UniSpeech, wav2vec2->unispeech, WAV_2_VEC_2->UNISPEECH class UniSpeechForSequenceClassification(UniSpeechPreTrainedModel): def __init__(self, config): super().__init__(config) self.unispeech = UniSpeechModel(config) num_layers = config.num_hidden_layers + 1 # transformer layers + input embeddings if config.use_weighted_layer_sum: self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers) self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size) self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels) # Initialize weights and apply final processing self.post_init() def freeze_feature_extractor(self): """ Calling this function will disable the gradient computation for the feature extractor so that its parameters will not be updated during training. """ self.unispeech.feature_extractor._freeze_parameters() def freeze_base_model(self): """ Calling this function will disable the gradient computation for the base model so that its parameters will not be updated during training. Only the classification head will be updated. """ for param in self.unispeech.parameters(): param.requires_grad = False @add_start_docstrings_to_model_forward(UNISPEECH_INPUTS_DOCSTRING) @add_code_sample_docstrings( processor_class=_SEQ_CLASS_PROCESSOR_FOR_DOC, checkpoint=_SEQ_CLASS_CHECKPOINT, output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC, modality="audio", ) def forward( self, input_values, attention_mask=None, output_attentions=None, output_hidden_states=None, return_dict=None, labels=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`): Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ..., config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss), If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict output_hidden_states = True if self.config.use_weighted_layer_sum else output_hidden_states outputs = self.unispeech( input_values, attention_mask=attention_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) if self.config.use_weighted_layer_sum: hidden_states = outputs[_HIDDEN_STATES_START_POSITION] hidden_states = torch.stack(hidden_states, dim=1) norm_weights = nn.functional.softmax(self.layer_weights, dim=-1) hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1) else: hidden_states = outputs[0] hidden_states = self.projector(hidden_states) if attention_mask is None: pooled_output = hidden_states.mean(dim=1) else: padding_mask = self._get_feature_vector_attention_mask(hidden_states.shape[1], attention_mask) hidden_states[~padding_mask] = 0.0 pooled_output = hidden_states.sum(dim=1) / padding_mask.sum(dim=1).view(-1, 1) logits = self.classifier(pooled_output) loss = None if labels is not None: loss_fct = CrossEntropyLoss() loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1)) if not return_dict: output = (logits,) + outputs[_HIDDEN_STATES_START_POSITION:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, )