751 lines
33 KiB
Python
751 lines
33 KiB
Python
import copy
|
|
import functools
|
|
import math
|
|
import random
|
|
from typing import Optional, Tuple
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch import distributed
|
|
from transformers.models.wav2vec2.modeling_wav2vec2 import _compute_mask_indices, _sample_negative_indices
|
|
from transformers.deepspeed import is_deepspeed_zero3_enabled
|
|
|
|
from models.arch_util import ResBlock
|
|
from trainer.networks import register_model
|
|
from utils.util import checkpoint
|
|
|
|
|
|
class Mel2Vec2FeatureProjection(nn.Module):
|
|
def __init__(self, inner_dim, dropout):
|
|
super().__init__()
|
|
self.layer_norm = nn.LayerNorm(inner_dim, eps=1e-5)
|
|
self.projection = nn.Linear(inner_dim, inner_dim)
|
|
self.dropout = nn.Dropout(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
|
|
|
|
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
|
|
|
|
|
|
class Wav2Vec2FeedForward(nn.Module):
|
|
def __init__(self, hidden_size, intermediate_size, dropout):
|
|
super().__init__()
|
|
self.intermediate_dropout = nn.Dropout(dropout)
|
|
|
|
self.intermediate_dense = nn.Linear(hidden_size, intermediate_size)
|
|
self.intermediate_act_fn = F.gelu
|
|
|
|
self.output_dense = nn.Linear(intermediate_size, hidden_size)
|
|
self.output_dropout = nn.Dropout(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, hidden_size, dropout):
|
|
super().__init__()
|
|
self.attention = Wav2Vec2Attention(
|
|
embed_dim=hidden_size,
|
|
num_heads=hidden_size//64,
|
|
dropout=dropout,
|
|
is_decoder=False,
|
|
)
|
|
self.dropout = nn.Dropout(dropout)
|
|
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-5)
|
|
self.feed_forward = Wav2Vec2FeedForward(hidden_size, hidden_size*2, dropout)
|
|
self.final_layer_norm = nn.LayerNorm(hidden_size, eps=1e-5)
|
|
|
|
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 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
|
|
|
|
|
|
from torch.nn.utils.weight_norm import WeightNorm
|
|
def __deepcopy__(self, memo):
|
|
# save and delete all weightnorm weights on self
|
|
weights = {}
|
|
for hook in self._forward_pre_hooks.values():
|
|
if isinstance(hook, WeightNorm):
|
|
weights[hook.name] = getattr(self, hook.name)
|
|
delattr(self, hook.name)
|
|
# remove this deepcopy method, restoring the object's original one if necessary
|
|
__deepcopy__ = self.__deepcopy__
|
|
if self.orig_deepcopy:
|
|
self.__deepcopy__ = self.orig_deepcopy
|
|
else:
|
|
del self.__deepcopy__
|
|
# actually do the copy
|
|
result = copy.deepcopy(self)
|
|
# restore weights and method on self
|
|
for name, value in weights.items():
|
|
setattr(self, name, value)
|
|
self.__deepcopy__ = __deepcopy__
|
|
return result
|
|
|
|
|
|
class Wav2Vec2PositionalConvEmbedding(nn.Module):
|
|
def __init__(self, hidden_size, num_conv_pos_embeddings=128, num_conv_pos_embedding_groups=16):
|
|
super().__init__()
|
|
self.conv = nn.Conv1d(
|
|
hidden_size,
|
|
hidden_size,
|
|
kernel_size=num_conv_pos_embeddings,
|
|
padding=num_conv_pos_embeddings // 2,
|
|
groups=num_conv_pos_embedding_groups,
|
|
)
|
|
# Fix weightnorm deepcopy; see: https://github.com/pytorch/pytorch/issues/28594
|
|
self.conv.orig_deepcopy = getattr(Wav2Vec2PositionalConvEmbedding, '__deepcopy__', None)
|
|
self.conv.__deepcopy__ = __deepcopy__.__get__(self.conv, self.conv.__class__)
|
|
|
|
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(num_conv_pos_embeddings)
|
|
self.activation = F.gelu
|
|
|
|
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 Wav2Vec2Encoder(nn.Module):
|
|
def __init__(self, hidden_size, dropout, num_layers, layerdrop):
|
|
super().__init__()
|
|
self.pos_conv_embed = Wav2Vec2PositionalConvEmbedding(hidden_size)
|
|
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-5)
|
|
self.dropout = nn.Dropout(dropout)
|
|
self.layers = nn.ModuleList([Wav2Vec2EncoderLayer(hidden_size, dropout) for _ in range(num_layers)])
|
|
self.layerdrop = layerdrop
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
attention_mask=None,
|
|
output_hidden_states=False,
|
|
):
|
|
all_hidden_states = () if output_hidden_states 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.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.layerdrop) else False
|
|
if not skip_the_layer or deepspeed_zero3_is_enabled:
|
|
layer_fn = functools.partial(layer, attention_mask=attention_mask)
|
|
layer_outputs = checkpoint(layer_fn, hidden_states)
|
|
hidden_states = layer_outputs[0]
|
|
|
|
hidden_states = self.layer_norm(hidden_states)
|
|
|
|
return hidden_states
|
|
|
|
|
|
class Mel2Vec(nn.Module):
|
|
def __init__(self,
|
|
mel_input_channels=256,
|
|
inner_dim=1024,
|
|
layers=24,
|
|
dropout=.1,
|
|
layerdrop=0,
|
|
mask_time_prob=.65,
|
|
mask_time_length=10,
|
|
disable_custom_linear_init=False,
|
|
linear_init_scale=.02,
|
|
feature_producer_type='standard',
|
|
):
|
|
super().__init__()
|
|
if feature_producer_type == 'standard':
|
|
self.input_blocks = nn.Sequential(nn.Conv1d(mel_input_channels, inner_dim//2, kernel_size=5, padding=2, stride=2),
|
|
nn.GroupNorm(num_groups=8, num_channels=inner_dim//2, affine=True),
|
|
nn.GELU(),
|
|
nn.Conv1d(inner_dim//2, inner_dim, kernel_size=3, padding=1, stride=2),
|
|
nn.GELU(),
|
|
nn.Conv1d(inner_dim, inner_dim, kernel_size=3, padding=1),
|
|
nn.GELU(),
|
|
)
|
|
self.dim_reduction_mult = 4
|
|
elif feature_producer_type == 'residual':
|
|
self.input_blocks = nn.Sequential(nn.Conv1d(mel_input_channels, inner_dim//2, kernel_size=5, padding=2, stride=2),
|
|
nn.GroupNorm(num_groups=8, num_channels=inner_dim//2, affine=True),
|
|
nn.GELU(),
|
|
ResBlock(dims=1, channels=inner_dim//2, dropout=dropout),
|
|
ResBlock(dims=1, channels=inner_dim//2, dropout=dropout),
|
|
nn.Conv1d(inner_dim//2, inner_dim, kernel_size=3, padding=1, stride=2),
|
|
nn.GELU(),
|
|
ResBlock(dims=1, channels=inner_dim, dropout=dropout),
|
|
ResBlock(dims=1, channels=inner_dim, dropout=dropout),
|
|
)
|
|
self.dim_reduction_mult = 4
|
|
elif feature_producer_type == 'voice_8x':
|
|
self.input_blocks = nn.Sequential(nn.Conv1d(mel_input_channels, inner_dim//4, kernel_size=5, padding=2, stride=2),
|
|
nn.GroupNorm(num_groups=8, num_channels=inner_dim//4, affine=True),
|
|
nn.GELU(),
|
|
nn.Conv1d(inner_dim//4, inner_dim//2, kernel_size=3, padding=1, stride=2),
|
|
nn.GELU(),
|
|
nn.Conv1d(inner_dim//2, inner_dim, kernel_size=3, padding=1, stride=2),
|
|
nn.GELU(),
|
|
nn.Conv1d(inner_dim, inner_dim, kernel_size=3, padding=1),
|
|
nn.GELU(),
|
|
)
|
|
self.dim_reduction_mult = 8
|
|
else:
|
|
assert False, f"feature_producer_type={feature_producer_type} not supported"
|
|
self.projector = Mel2Vec2FeatureProjection(inner_dim, dropout)
|
|
self.masked_spec_embed = nn.Parameter(torch.rand(inner_dim,))
|
|
self.encoder = Wav2Vec2Encoder(inner_dim, dropout, layers, layerdrop)
|
|
self.mask_time_prob = mask_time_prob
|
|
self.mask_time_length = mask_time_length
|
|
self.disable_custom_linear_init = disable_custom_linear_init
|
|
self.linear_init_scale = linear_init_scale
|
|
self.mel_dim = mel_input_channels
|
|
self.apply(self.init)
|
|
|
|
def init(self, module):
|
|
"""Initialize the weights"""
|
|
# gumbel softmax requires special init
|
|
if isinstance(module, Wav2Vec2PositionalConvEmbedding):
|
|
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, Mel2Vec2FeatureProjection):
|
|
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):
|
|
if self.disable_custom_linear_init:
|
|
return
|
|
module.weight.data.normal_(mean=0.0, std=self.linear_init_scale)
|
|
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 apply_masking(
|
|
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](https://arxiv.org/abs/1904.08779).
|
|
"""
|
|
|
|
# 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.mask_time_prob > 0 and self.training:
|
|
mask_time_indices = _compute_mask_indices(
|
|
(batch_size, sequence_length),
|
|
mask_prob=self.mask_time_prob,
|
|
mask_length=self.mask_time_length,
|
|
attention_mask=attention_mask,
|
|
min_masks=self.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)
|
|
|
|
return hidden_states
|
|
|
|
def forward(self, mel, mask_time_indices=None, return_projections=False):
|
|
proj = self.input_blocks(mel).permute(0,2,1)
|
|
proj, norm_proj = self.projector(proj)
|
|
|
|
# Mask projections
|
|
h = self.apply_masking(proj, mask_time_indices)
|
|
h = self.encoder(h)
|
|
|
|
if return_projections:
|
|
return h, norm_proj
|
|
return h
|
|
|
|
|
|
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, proj_dim=1024, codevector_dim=512, num_codevector_groups=2, num_codevectors_per_group=320):
|
|
super().__init__()
|
|
self.codevector_dim = codevector_dim
|
|
self.num_groups = num_codevector_groups
|
|
self.num_vars = num_codevectors_per_group
|
|
self.num_codevectors = num_codevector_groups * num_codevectors_per_group
|
|
|
|
if codevector_dim % self.num_groups != 0:
|
|
raise ValueError(
|
|
f"`codevector_dim {codevector_dim} must be divisible "
|
|
f"by `num_codevector_groups` {num_codevector_groups} for concatenation"
|
|
)
|
|
|
|
# storage for codebook variables (codewords)
|
|
self.codevectors = nn.Parameter(
|
|
torch.FloatTensor(1, self.num_groups * self.num_vars, codevector_dim // self.num_groups)
|
|
)
|
|
self.weight_proj = nn.Linear(proj_dim, self.num_groups * self.num_vars)
|
|
|
|
# can be decayed for training
|
|
self.temperature = 2
|
|
|
|
# Parameters init.
|
|
self.weight_proj.weight.data.normal_(mean=0.0, std=1)
|
|
self.weight_proj.bias.data.zero_()
|
|
nn.init.uniform_(self.codevectors)
|
|
|
|
@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 get_codes(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)
|
|
codevector_idx = hidden_states.argmax(dim=-1)
|
|
idxs = codevector_idx.view(batch_size, sequence_length, self.num_groups)
|
|
return idxs
|
|
|
|
|
|
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 differentiable 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
|
|
# compute 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 ContrastiveTrainingWrapper(nn.Module):
|
|
def __init__(self, inner_dim=1024, dropout=.1, mask_time_prob=.65, mask_time_length=6, num_negatives=100,
|
|
max_gumbel_temperature=2.0, min_gumbel_temperature=.5, gumbel_temperature_decay=.999995,
|
|
codebook_size=320, codebook_groups=2, freq_mask_percent=0, inp_length_multiplier=256,
|
|
do_reconstruction_loss=False,
|
|
**kwargs):
|
|
super().__init__()
|
|
self.m2v = Mel2Vec(inner_dim=inner_dim, dropout=dropout, mask_time_prob=mask_time_prob,
|
|
mask_time_length=mask_time_length, **kwargs)
|
|
self.num_negatives = num_negatives
|
|
self.mask_time_prob = mask_time_prob
|
|
self.mask_time_length = mask_time_length
|
|
self.max_gumbel_temperature = max_gumbel_temperature
|
|
self.min_gumbel_temperature = min_gumbel_temperature
|
|
self.gumbel_temperature_decay = gumbel_temperature_decay
|
|
self.freq_mask_percent = freq_mask_percent
|
|
self.quantizer = Wav2Vec2GumbelVectorQuantizer(inner_dim, num_codevector_groups=codebook_groups, num_codevectors_per_group=codebook_size)
|
|
self.num_losses_record = []
|
|
self.inp_length_factor = inp_length_multiplier
|
|
|
|
# make sure that project_hid & project_q are initialized like normal linear layers
|
|
self.project_hid = nn.Linear(inner_dim, self.quantizer.codevector_dim)
|
|
self.project_q = nn.Linear(self.quantizer.codevector_dim, self.quantizer.codevector_dim)
|
|
|
|
self.reconstruction = do_reconstruction_loss
|
|
if do_reconstruction_loss:
|
|
blocks = [[ResBlock(dims=1, channels=inner_dim, dropout=dropout),
|
|
ResBlock(dims=1, channels=inner_dim, dropout=dropout, use_conv=True, up=True)] for _ in range(int(math.log2(self.m2v.dim_reduction_mult)))]
|
|
blocks = sum(blocks, [])
|
|
blocks.append(nn.Conv1d(inner_dim, self.m2v.mel_dim, kernel_size=3, padding=1))
|
|
self.reconstruction_net = nn.Sequential(
|
|
nn.Conv1d(self.quantizer.codevector_dim, inner_dim, kernel_size=3, padding=1),
|
|
*blocks
|
|
)
|
|
|
|
@staticmethod
|
|
def compute_contrastive_logits(
|
|
target_features: torch.FloatTensor,
|
|
negative_features: torch.FloatTensor,
|
|
predicted_features: torch.FloatTensor,
|
|
temperature: int = 0.1,
|
|
):
|
|
"""
|
|
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
|
|
|
|
def update_for_step(self, step, *args):
|
|
self.quantizer.temperature = max(
|
|
self.max_gumbel_temperature * self.gumbel_temperature_decay**step,
|
|
self.min_gumbel_temperature,
|
|
)
|
|
|
|
def get_grad_norm_parameter_groups(self):
|
|
groups = {
|
|
'projector': list(self.m2v.input_blocks.parameters()) + list(self.m2v.projector.parameters()),
|
|
'encoder': list(self.m2v.encoder.parameters()),
|
|
'output_blocks': list(self.project_hid.parameters()) + list(self.project_q.parameters()),
|
|
}
|
|
return groups
|
|
|
|
def get_codes(self, mel):
|
|
proj = self.m2v.input_blocks(mel).permute(0,2,1)
|
|
_, proj = self.m2v.projector(proj)
|
|
codes = self.quantizer.get_codes(proj)
|
|
return codes
|
|
|
|
def forward(self, mel, inp_lengths=None):
|
|
mel = mel[:, :, :-1] # The MEL computation always pads with 1, throwing off optimal tensor math.
|
|
features_shape = (mel.shape[0], mel.shape[-1]//self.m2v.dim_reduction_mult)
|
|
|
|
# Frequency masking
|
|
freq_mask_width = int(random.random() * self.freq_mask_percent * mel.shape[1])
|
|
if freq_mask_width >= 2:
|
|
freq_start = random.randint(0, mel.shape[1]-freq_mask_width)
|
|
mel[:, freq_start:freq_start+freq_mask_width] = 0
|
|
|
|
# Build input masks from inp_lengths if possible.
|
|
attention_mask = torch.ones(features_shape, device=mel.device, dtype=torch.long)
|
|
if inp_lengths is not None:
|
|
inp_lengths = inp_lengths // (self.inp_length_factor*self.m2v.dim_reduction_mult)
|
|
for i, l in enumerate(inp_lengths):
|
|
attention_mask[i, l:] = 0
|
|
|
|
mask_time_indices = _compute_mask_indices(features_shape, self.mask_time_prob, self.mask_time_length, attention_mask=attention_mask)
|
|
sampled_negative_indices = torch.tensor(_sample_negative_indices(features_shape, self.num_negatives, mask_time_indices=mask_time_indices), device=mel.device)
|
|
mask_time_indices = torch.tensor(mask_time_indices, device=mel.device)
|
|
|
|
outputs, proj = self.m2v(mel, mask_time_indices, return_projections=True)
|
|
|
|
# 1. project all transformed features (including masked) to final vq dim
|
|
transformer_features = self.project_hid(outputs)
|
|
|
|
# 2. quantize all (unmasked) extracted features and project to final vq dim
|
|
quantized_features, codevector_perplexity = self.quantizer(
|
|
proj, mask_time_indices=mask_time_indices
|
|
)
|
|
quantized_features = self.project_q(quantized_features)
|
|
batch_size, sequence_length, hidden_size = quantized_features.shape
|
|
|
|
# 3. sample K negatives (distractors) quantized states for contrastive loss
|
|
# if attention_mask is passed, make sure that padded feature vectors cannot be sampled
|
|
# sample negative quantized vectors BTC => (BxT)C
|
|
negative_quantized_features = quantized_features.view(-1, hidden_size)[
|
|
sampled_negative_indices.long().view(-1)
|
|
]
|
|
negative_quantized_features = negative_quantized_features.view(
|
|
batch_size, sequence_length, -1, hidden_size
|
|
).permute(2, 0, 1, 3)
|
|
|
|
# 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,
|
|
.1,
|
|
)
|
|
|
|
# 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))
|
|
logits = 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(logits.float(), target, reduction="mean")
|
|
# 7. compute diversity loss: \mathbf{L}_d
|
|
num_codevectors = self.quantizer.num_codevectors
|
|
diversity_loss = (num_codevectors - codevector_perplexity) / num_codevectors
|
|
|
|
if self.reconstruction:
|
|
reconstruction = self.reconstruction_net(quantized_features.permute(0,2,1))
|
|
reconstruction_loss = F.mse_loss(reconstruction, mel)
|
|
return contrastive_loss, diversity_loss, reconstruction_loss
|
|
|
|
return contrastive_loss, diversity_loss
|
|
|
|
|
|
@register_model
|
|
def register_mel2vec_pretraining(opt_net, opt):
|
|
return ContrastiveTrainingWrapper(**opt_net['kwargs'])
|
|
|
|
|
|
@register_model
|
|
def register_mel2vec(opt_net, opt):
|
|
return Mel2Vec(**opt_net['kwargs'])
|
|
|
|
|
|
if __name__ == '__main__':
|
|
model = ContrastiveTrainingWrapper(freq_mask_percent=.5, do_reconstruction_loss=True, feature_producer_type='residual')
|
|
mel = torch.randn((2,256,401))
|
|
print(model(mel)) |