# 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 Wav2Vec2 model. """
import warnings
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 transformers.deepspeed import is_deepspeed_zero3_enabled
from ...activations import ACT2FN
from ...file_utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
replace_return_docstrings,
)
from ...modeling_outputs import BaseModelOutput, CausalLMOutput, MaskedLMOutput
from ...modeling_utils import PreTrainedModel
from ...utils import logging
from .configuration_wav2vec2 import Wav2Vec2Config
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "Wav2Vec2Config"
WAV_2_VEC_2_PRETRAINED_MODEL_ARCHIVE_LIST = [
"facebook/wav2vec2-base-960h",
"facebook/wav2vec2-large-960h",
"facebook/wav2vec2-large-960h-lv60",
"facebook/wav2vec2-large-960h-lv60-self",
# See all Wav2Vec2 models at https://huggingface.co/models?filter=wav2vec2
]
@dataclass
class Wav2Vec2BaseModelOutput(ModelOutput):
"""
Output type of :class:`~transformers.Wav2Vec2BaseModelOutput`, 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 Wav2Vec2ForPreTrainingOutput(ModelOutput):
"""
Output type of :class:`~transformers.Wav2Vec2ForPreTrainingOutput`, 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 <https://arxiv.org/pdf/2006.11477.pdf>`__ . (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
def _compute_mask_indices(
shape: Tuple[int, int],
mask_prob: float,
mask_length: int,
device: torch.device,
min_masks: int = 0,
) -> torch.tensor:
"""
Computes random mask spans for a given shape. Used to implement `SpecAugment: A Simple Data Augmentation Method for
ASR <https://arxiv.org/abs/1904.08779>`__.
Args:
shape: the the shape for which to compute masks.
should be of size 2 where first element is batch size and 2nd is timesteps
mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
however due to overlaps, the actual number will be smaller (unless no_overlap is True)
mask_length: size of the mask
min_masks: minimum number of masked spans
"""
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} and `sequence_length`: {sequence_length}`"
)
# compute number of masked spans in batch
num_masked_spans = int(mask_prob * sequence_length / mask_length + torch.rand((1,)).item())
num_masked_spans = max(num_masked_spans, min_masks)
# make sure num masked indices <= sequence_length
if num_masked_spans * mask_length > sequence_length:
num_masked_spans = sequence_length // mask_length
# SpecAugment mask to fill
spec_aug_mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
# uniform distribution to sample from, make sure that offset samples are < sequence_length
uniform_dist = torch.ones((batch_size, sequence_length - (mask_length - 1)), device=device)
# get random indices to mask
spec_aug_mask_idxs = torch.multinomial(uniform_dist, num_masked_spans)
# expand masked indices to masked spans
spec_aug_mask_idxs = (
spec_aug_mask_idxs.unsqueeze(dim=-1)
.expand((batch_size, num_masked_spans, mask_length))
.reshape(batch_size, num_masked_spans * mask_length)
)
offsets = (
torch.arange(mask_length, device=device)[None, None, :]
.expand((batch_size, num_masked_spans, mask_length))
.reshape(batch_size, num_masked_spans * mask_length)
)
spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
# scatter indices to mask
spec_aug_mask = spec_aug_mask.scatter(1, spec_aug_mask_idxs, True)
return spec_aug_mask
class Wav2Vec2NoLayerNormConvLayer(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
self.in_conv_dim = config.conv_dim[layer_id] 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
class Wav2Vec2LayerNormConvLayer(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
self.in_conv_dim = config.conv_dim[layer_id] 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
class Wav2Vec2GroupNormConvLayer(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
self.in_conv_dim = config.conv_dim[layer_id] 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
class Wav2Vec2PositionalConvEmbedding(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 = Wav2Vec2SamePadLayer(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
class Wav2Vec2SamePadLayer(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
class Wav2Vec2FeatureExtractor(nn.Module):
"""Construct the featurs from raw audio waveform"""
def __init__(self, config):
super().__init__()
if config.feat_extract_norm == "group":
conv_layers = [Wav2Vec2GroupNormConvLayer(config, layer_id=0)] + [
Wav2Vec2NoLayerNormConvLayer(config, layer_id=i + 1) for i in range(config.num_feat_extract_layers - 1)
]
elif config.feat_extract_norm == "layer":
conv_layers = [
Wav2Vec2LayerNormConvLayer(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)
def _freeze_parameters(self):
for param in self.parameters():
param.requires_grad = False
def forward(self, input_values):
hidden_states = input_values[:, None]
for conv_layer in self.conv_layers:
hidden_states = conv_layer(hidden_states)
return hidden_states
class Wav2Vec2FeatureProjection(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->Wav2Vec2
class Wav2Vec2Attention(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
assert (
self.head_dim * num_heads == self.embed_dim
), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} 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, embed_dim = 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)
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped, past_key_value
class Wav2Vec2FeedForward(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
class Wav2Vec2EncoderLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.attention = Wav2Vec2Attention(
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 = Wav2Vec2FeedForward(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
class Wav2Vec2EncoderLayerStableLayerNorm(nn.Module):
def __init__(self, config):
super().__init__()
self.attention = Wav2Vec2Attention(
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 = Wav2Vec2FeedForward(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
class Wav2Vec2Encoder(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.pos_conv_embed = Wav2Vec2PositionalConvEmbedding(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([Wav2Vec2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
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 getattr(self.config, "gradient_checkpointing", False) 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,
)
class Wav2Vec2EncoderStableLayerNorm(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.pos_conv_embed = Wav2Vec2PositionalConvEmbedding(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(
[Wav2Vec2EncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
)
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 getattr(self.config, "gradient_checkpointing", False) 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 Wav2Vec2GumbelVectorQuantizer(nn.Module):
"""
Vector quantization using gumbel softmax. See `CATEGORICAL REPARAMETERIZATION WITH GUMBEL-SOFTMAX
<https://arxiv.org/pdf/1611.01144.pdf>`__ 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 = 1
def set_temperature(self, temperature: int):
self.temperature = temperature
@staticmethod
def _compute_perplexity(probs, mask=None):
if mask is not None:
mask_extended = mask.flatten()[:, None, None].expand(probs.shape)
probs = torch.where(mask_extended, probs, torch.zeros_like(probs))
marginal_probs = probs.sum(dim=0) / mask.sum()
else:
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, mask_time_indices=None):
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, mask_time_indices)
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, mask_time_indices)
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)
.sum(-2)
.view(batch_size, sequence_length, -1)
)
return codevectors, perplexity
class Wav2Vec2PreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = Wav2Vec2Config
base_model_prefix = "wav2vec2"
_keys_to_ignore_on_load_missing = [r"position_ids"]
def _init_weights(self, module):
"""Initialize the weights"""
# gumbel softmax requires special init
if isinstance(module, Wav2Vec2GumbelVectorQuantizer):
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, nn.Linear):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
elif isinstance(module, nn.Conv1d):
if is_deepspeed_zero3_enabled():
import deepspeed
if hasattr(module, "weight_v") and hasattr(module, "weight_g"):
with deepspeed.zero.GatheredParameters([module.weight_v, module.weight_g], modifier_rank=0):
nn.init.kaiming_normal_(module.weight.data)
else:
with deepspeed.zero.GatheredParameters(module.weight, modifier_rank=0):
nn.init.kaiming_normal_(module.weight.data)
else:
nn.init.kaiming_normal_(module.weight.data)
if isinstance(module, (nn.Linear, nn.Conv1d)) and module.bias is not None:
module.bias.data.zero_()
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
WAV_2_VEC_2_START_DOCSTRING = r"""
Wav2Vec2 was proposed in `wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
<https://arxiv.org/abs/2006.11477>`__ by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
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 <https://pytorch.org/docs/stable/nn.html#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.Wav2Vec2Config`): 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.
"""
WAV_2_VEC_2_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.Wav2Vec2Processor` should
be used for padding and conversion into a tensor of type `torch.FloatTensor`. See
:meth:`transformers.Wav2Vec2Processor.__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``, such as `wav2vec2-base
<https://huggingface.co/facebook/wav2vec2-base-960h>`__, :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.
"""
[docs]@add_start_docstrings(
"The bare Wav2Vec2 Model transformer outputting raw hidden-states without any specific head on top.",
WAV_2_VEC_2_START_DOCSTRING,
)
class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
def __init__(self, config: Wav2Vec2Config):
super().__init__(config)
self.config = config
self.feature_extractor = Wav2Vec2FeatureExtractor(config)
self.feature_projection = Wav2Vec2FeatureProjection(config)
self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
if config.do_stable_layer_norm:
self.encoder = Wav2Vec2EncoderStableLayerNorm(config)
else:
self.encoder = Wav2Vec2Encoder(config)
self.init_weights()
def _mask_hidden_states(
self, hidden_states: torch.FloatTensor, mask_time_indices: Optional[torch.FloatTensor] = None
):
"""
Masks extracted features along time axis and/or along feature axis according to `SpecAugment
<https://arxiv.org/abs/1904.08779>`__ .
"""
# `config.apply_spec_augment` can set masking to False
if not getattr(self.config, "apply_spec_augment", True):
return hidden_states
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:
# generate indices & apply SpecAugment along time axis
batch_size, sequence_length, hidden_size = hidden_states.size()
mask_time_indices = _compute_mask_indices(
(batch_size, sequence_length),
mask_prob=self.config.mask_time_prob,
mask_length=self.config.mask_time_length,
device=hidden_states.device,
min_masks=2,
)
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,
device=hidden_states.device,
)
hidden_states[mask_feature_indices[:, None].expand(-1, sequence_length, -1)] = 0
return hidden_states
[docs] @add_start_docstrings_to_model_forward(WAV_2_VEC_2_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_values,
attention_mask=None,
mask_time_indices=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
"""
Returns:
Example::
>>> from transformers import Wav2Vec2Processor, Wav2Vec2Model
>>> from datasets import load_dataset
>>> import soundfile as sf
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
>>> model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")
>>> 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 = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
>>> hidden_states = model(input_values).last_hidden_state
"""
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 real output lengths according to convolution formula
output_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long)
attention_mask = torch.zeros(
extract_features.shape[:2], dtype=extract_features.dtype, device=extract_features.device
)
# these two operations makes sure that all values
# before the output lengths indices are attended to
attention_mask[
(torch.arange(attention_mask.shape[0], device=extract_features.device), output_lengths - 1)
] = 1
attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
hidden_states, extract_features = self.feature_projection(extract_features)
if mask_time_indices is not None: # apply SpecAugment along time axis with given indices
hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
hidden_states = self._mask_hidden_states(hidden_states)
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 Wav2Vec2BaseModelOutput(
last_hidden_state=hidden_states,
extract_features=extract_features,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
[docs]@add_start_docstrings("""Wav2Vec2 Model with a quantizer and `VQ` head on top. """, WAV_2_VEC_2_START_DOCSTRING)
class Wav2Vec2ForPreTraining(Wav2Vec2PreTrainedModel):
def __init__(self, config: Wav2Vec2Config):
super().__init__(config)
self.wav2vec2 = Wav2Vec2Model(config)
self.dropout_features = nn.Dropout(config.feat_quantizer_dropout)
self.quantizer = Wav2Vec2GumbelVectorQuantizer(config)
self.project_q = nn.Linear(config.codevector_dim, config.proj_codevector_dim)
self.project_hid = nn.Linear(config.hidden_size, config.proj_codevector_dim)
self.init_weights()
def set_gumbel_temperature(self, temperature: int):
"""
Set the Gumbel softmax temperature to a given value. Only necessary for training
"""
return self.quantizer.set_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.wav2vec2.feature_extractor._freeze_parameters()
@staticmethod
def _sample_negatives(features: torch.FloatTensor, num_negatives: int):
"""
Sample `num_negatives` vectors from feature vectors.
"""
batch_size, sequence_length, hidden_size = features.shape
if sequence_length <= 1:
raise ValueError(
f"`features should have `sequence_length` > 1, but are of shape (batch_size, sequence_length, hidden_size) = ({batch_size, sequence_length, hidden_size})."
)
features = features.view(-1, hidden_size) # BTC => (BxT)C
with torch.no_grad():
# get `num_negatives` random vector indices from the same utterance
sampled_negative_indices = torch.randint(
low=0,
high=sequence_length - 1,
size=(batch_size, num_negatives * sequence_length),
device=features.device,
)
# generate indices of the positive vectors themselves, repeat them `num_negatives` times
feature_indices = (
torch.arange(sequence_length, device=features.device)[:, None]
.expand(sequence_length, num_negatives)
.flatten()
)
# avoid sampling the same positive vector, but keep the distribution uniform
sampled_negative_indices[sampled_negative_indices >= feature_indices] += 1
# correct for batch size
for batch_idx in range(1, batch_size):
sampled_negative_indices[batch_idx] += batch_idx * sequence_length
# take negative vectors from sampled indices
sampled_negatives = features[sampled_negative_indices.view(-1)]
sampled_negatives = sampled_negatives.view(batch_size, sequence_length, num_negatives, hidden_size).permute(
2, 0, 1, 3
)
return sampled_negatives
@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
`[positive_feature, negative_features]` and `[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).type_as(
target_features
)
# apply temperature
logits = logits / temperature
return logits
[docs] @add_start_docstrings_to_model_forward(WAV_2_VEC_2_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_values,
attention_mask=None,
mask_time_indices=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.
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
if mask_time_indices is not None:
mask_time_indices = mask_time_indices.to(torch.bool)
outputs = self.wav2vec2(
input_values,
attention_mask=attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
mask_time_indices=mask_time_indices,
return_dict=return_dict,
)
# 1. project all transformed features (including masked) to final vq dim
transformer_features = self.project_hid(outputs[0])
# 2. 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, mask_time_indices)
quantized_features = self.project_q(quantized_features)
loss = None
if self.training:
# for training, we sample negatives
# 3. sample K negatives (distractors) quantized states for contrastive loss
negative_quantized_features = self._sample_negatives(quantized_features, self.config.num_negatives)
# 4. compute logits, corresponding to `logs = sim(c_t, [q_t, \sim{q}_t]) / \kappa`
# of equation (3) in https://arxiv.org/pdf/2006.11477.pdf
logits = self.compute_contrastive_logits(
quantized_features[None, :],
negative_quantized_features,
transformer_features,
self.config.contrastive_logits_temperature,
)
# 5. if a negative vector is identical to the positive (i.e. when codebook utilization is low),
# its cosine similarity will be masked
neg_is_pos = (quantized_features == negative_quantized_features).all(-1)
if neg_is_pos.any():
logits[1:][neg_is_pos] = float("-inf")
# 6. compute contrastive loss \mathbf{L}_m = cross_entropy(logs) =
# -log(exp(sim(c_t, q_t)/\kappa) / \sum_{\sim{q}} exp(sim(c_t, \sim{q})/\kappa))
preds = logits.transpose(0, 2).reshape(-1, logits.size(0))
target = ((1 - mask_time_indices.long()) * -100).transpose(0, 1).flatten()
contrastive_loss = nn.functional.cross_entropy(preds.float(), target, reduction="sum")
# 7. compute diversity loss: \mathbf{L}_d
num_codevectors = self.config.num_codevectors_per_group * self.config.num_codevector_groups
diversity_loss = (num_codevectors - codevector_perplexity) / num_codevectors
# 8. \mathbf{L} = \mathbf{L}_m + \alpha * \mathbf{L}_d
loss = contrastive_loss + self.config.diversity_loss_weight * diversity_loss
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 Wav2Vec2ForPreTrainingOutput(
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("""Wav2Vec2 Model with a `language modeling` head on top. """, WAV_2_VEC_2_START_DOCSTRING)
class Wav2Vec2ForMaskedLM(Wav2Vec2PreTrainedModel):
def __init__(self, config):
super().__init__(config)
warnings.warn(
"The class `Wav2Vec2ForMaskedLM` is deprecated. Please use `Wav2Vec2ForCTC` instead.", FutureWarning
)
self.wav2vec2 = Wav2Vec2Model(config)
self.dropout = nn.Dropout(config.final_dropout)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)
self.init_weights()
@add_start_docstrings_to_model_forward(WAV_2_VEC_2_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutput, 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, sequence_length, hidden_size)`, `optional`):
TODO(PVP): Fill out when adding training
Returns:
Example::
>>> from transformers import Wav2Vec2Processor, Wav2Vec2Model
>>> from datasets import load_dataset
>>> import soundfile as sf
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
>>> model = Wav2Vec2ForMaskedLM.from_pretrained("facebook/wav2vec2-base-960h")
>>> 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 = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
>>> logits = model(input_values).logits
>>> predicted_ids = torch.argmax(logits, dim=-1)
>>> transcription = processor.decode(predicted_ids[0])
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.wav2vec2(
input_values,
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)
if not return_dict:
output = (logits,) + outputs[2:]
return output
return MaskedLMOutput(logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)
[docs]@add_start_docstrings(
"""Wav2Vec2 Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC). """,
WAV_2_VEC_2_START_DOCSTRING,
)
class Wav2Vec2ForCTC(Wav2Vec2PreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.wav2vec2 = Wav2Vec2Model(config)
self.dropout = nn.Dropout(config.final_dropout)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size)
self.init_weights()
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.wav2vec2.feature_extractor._freeze_parameters()
[docs] @add_start_docstrings_to_model_forward(WAV_2_VEC_2_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=BaseModelOutput, 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]``.
Returns:
Example::
>>> import torch
>>> from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
>>> from datasets import load_dataset
>>> import soundfile as sf
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
>>> model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
>>> 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 = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
>>> logits = model(input_values).logits
>>> predicted_ids = torch.argmax(logits, dim=-1)
>>> transcription = processor.decode(predicted_ids[0])
>>> # compute loss
>>> target_transcription = "A MAN SAID TO THE UNIVERSE SIR I EXIST"
>>> # wrap processor as target processor to encode labels
>>> with processor.as_target_processor():
>>> labels = processor(target_transcription, return_tensors="pt").input_ids
>>> loss = model(input_values, labels=labels).loss
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.wav2vec2(
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:
# 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[2:]
return ((loss,) + output) if loss is not None else output
return CausalLMOutput(
loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions
)