314 lines
16 KiB
Python
314 lines
16 KiB
Python
import functools
|
|
from math import log
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from transformers import GPT2Model, GPT2Config
|
|
|
|
from models.arch_util import AttentionBlock
|
|
from models.gpt_voice.gpt_asr_hf import GPT2InferenceModel
|
|
from models.tacotron2.text import symbols
|
|
from trainer.networks import register_model
|
|
from utils.util import opt_get
|
|
|
|
|
|
def null_position_embeddings(range, dim):
|
|
return torch.zeros((range.shape[0], range.shape[1], dim), device=range.device)
|
|
|
|
|
|
class ConditioningEncoder(nn.Module):
|
|
def __init__(self,
|
|
spec_dim,
|
|
embedding_dim,
|
|
attn_blocks=6,
|
|
num_attn_heads=4,
|
|
do_checkpointing=False):
|
|
super().__init__()
|
|
attn = []
|
|
self.init = nn.Conv1d(spec_dim, embedding_dim, kernel_size=1)
|
|
for a in range(attn_blocks):
|
|
attn.append(AttentionBlock(embedding_dim, num_attn_heads, do_checkpoint=do_checkpointing))
|
|
self.attn = nn.Sequential(*attn)
|
|
self.dim = embedding_dim
|
|
self.do_checkpointing = do_checkpointing
|
|
|
|
def forward(self, x):
|
|
h = self.init(x)
|
|
h = self.attn(h)
|
|
return h[:, :, 0]
|
|
|
|
|
|
class TopEncoder(nn.Module):
|
|
def __init__(self, layers, dim, heads, do_checkpointing=False, dim_reduction=16):
|
|
self.init = nn.Conv1d(dim, dim, kernel_size=1)
|
|
reduction_layers = []
|
|
for j in range(int(log(dim_reduction, 2))):
|
|
reduction_layers.append(AttentionBlock(dim, heads, do_checkpoint=do_checkpointing))
|
|
reduction_layers.append(nn.Conv1d(dim, dim, kernel_size=3, padding=1, stride=2))
|
|
self.reduction_layers = nn.Sequential(*reduction_layers)
|
|
actual_layers = [AttentionBlock(dim, heads, do_checkpoint=do_checkpointing) for _ in range(layers)]
|
|
self.actual_layers = nn.Sequential(*actual_layers)
|
|
|
|
def forward(self, x):
|
|
h = self.init(x)
|
|
h = self.reduction_layers(h)
|
|
h = self.actual_layers(h)
|
|
return h
|
|
|
|
|
|
class UnifiedGptVoice(nn.Module):
|
|
"""
|
|
Derived from GptTtsHf, but offers multiple modes of autoregressive operation:
|
|
- Text only
|
|
- Voice only
|
|
- Text conditioned on voice
|
|
- Voice conditioned on text
|
|
"""
|
|
|
|
def __init__(self, top_encoder_layers=4, top_layers=8, bottom_layers=8, top_dim_reduction=16, model_dim=512, heads=8,
|
|
max_symbols_per_phrase=120, max_mel_tokens=250, max_total_tokens=370, max_conditioning_inputs=3,
|
|
checkpointing=True, mel_length_compression=1024, max_conditioning_length=60, number_text_tokens=256,
|
|
start_text_token=255, stop_text_token=0, number_mel_codes=8194, start_mel_token=8192,
|
|
stop_mel_token=8193):
|
|
super().__init__()
|
|
|
|
self.number_text_tokens = number_text_tokens
|
|
self.start_text_token = start_text_token
|
|
self.stop_text_token = stop_text_token
|
|
self.number_mel_codes = number_mel_codes
|
|
self.start_mel_token = start_mel_token
|
|
self.stop_mel_token = stop_mel_token
|
|
|
|
self.max_mel_tokens = max_mel_tokens
|
|
self.max_symbols_per_phrase = max_symbols_per_phrase
|
|
self.max_total_tokens = max_total_tokens
|
|
self.model_dim = model_dim
|
|
self.max_conditioning_inputs = max_conditioning_inputs
|
|
self.mel_length_compression = mel_length_compression
|
|
self.conditioning_encoder = ConditioningEncoder(80, model_dim, num_attn_heads=heads)
|
|
self.text_embedding = nn.Embedding(self.number_text_tokens, model_dim)
|
|
self.text_pos_solo_embedding = nn.Embedding(self.max_symbols_per_phrase + 1, model_dim)
|
|
self.text_pos_paired_embedding = nn.Embedding(self.max_symbols_per_phrase + 1, model_dim)
|
|
self.mel_pos_solo_embedding = nn.Embedding(self.max_mel_tokens + 1, model_dim)
|
|
self.mel_pos_paired_embedding = nn.Embedding(self.max_mel_tokens + 1, model_dim)
|
|
seq_length = 2+self.max_total_tokens+self.max_conditioning_inputs
|
|
|
|
self.top_encoder = TopEncoder(top_encoder_layers, model_dim, heads, do_checkpointing=checkpointing,
|
|
dim_reduction=top_dim_reduction)
|
|
self.top_gpt_config = GPT2Config(vocab_size=1,
|
|
n_positions=seq_length // top_dim_reduction,
|
|
n_ctx=seq_length // top_dim_reduction,
|
|
n_embd=model_dim,
|
|
n_layer=top_layers,
|
|
n_head=heads,
|
|
gradient_checkpointing=checkpointing,
|
|
use_cache=not checkpointing)
|
|
self.top_gpt = GPT2Model(self.top_gpt_config)
|
|
del self.top_gpt.wte
|
|
self.top_gpt_start_embedding = nn.Parameter(torch.randn(1,1,model_dim)*self.top_gpt_config.initializer_range,
|
|
requires_grad=True)
|
|
self.top_dim_reduction = top_dim_reduction
|
|
|
|
self.bottom_gpt_config = GPT2Config(vocab_size=self.number_mel_codes,
|
|
n_positions=seq_length,
|
|
n_ctx=seq_length,
|
|
n_embd=model_dim,
|
|
n_layer=bottom_layers,
|
|
n_head=heads,
|
|
gradient_checkpointing=checkpointing,
|
|
use_cache=not checkpointing)
|
|
self.bottom_gpt = GPT2Model(self.bottom_gpt_config)
|
|
# Override the built in positional embeddings
|
|
del self.bottom_gpt.wpe
|
|
self.bottom_gpt.wpe = functools.partial(null_position_embeddings, dim=model_dim)
|
|
|
|
self.final_norm = nn.LayerNorm(model_dim)
|
|
self.text_head = nn.Linear(model_dim, self.number_text_tokens)
|
|
self.mel_head = nn.Linear(model_dim, self.number_mel_codes)
|
|
self.max_conditioning_length = max_conditioning_length
|
|
|
|
# Initialize the embeddings per the GPT-2 scheme
|
|
for module in [self.text_embedding, self.text_pos_solo_embedding, self.text_pos_paired_embedding,
|
|
self.mel_pos_solo_embedding, self.mel_pos_paired_embedding]:
|
|
module.weight.data.normal_(mean=0.0, std=self.bottom_gpt.config.initializer_range)
|
|
if module.padding_idx is not None:
|
|
module.weight.data[module.padding_idx].zero_()
|
|
|
|
def build_aligned_inputs_and_targets(self, input, start_token, stop_token):
|
|
inp = F.pad(input, (1,0), value=start_token)
|
|
tar = F.pad(input, (0,1), value=stop_token)
|
|
return inp, tar
|
|
|
|
def set_mel_padding(self, mel_input_tokens, wav_lengths):
|
|
"""
|
|
Given mel tokens that are derived from a padded audio clip and the actual lengths of each batch element in
|
|
that audio clip, reformats the tokens with STOP_MEL_TOKEN in place of the zero padding. This is required
|
|
preformatting to create a working TTS model.
|
|
"""
|
|
# Set padding areas within MEL (currently it is coded with the MEL code for <zero>).
|
|
mel_lengths = wav_lengths // self.mel_length_compression
|
|
for b in range(len(mel_lengths)):
|
|
actual_end = mel_lengths[b] + 1 # Due to the convolutional nature of how these tokens are generated, it would be best if the model predicts a token past the actual last token.
|
|
if actual_end < mel_input_tokens.shape[-1]:
|
|
mel_input_tokens[b, actual_end:] = self.stop_mel_token
|
|
return mel_input_tokens
|
|
|
|
def randomly_permute_conditioning_input(self, speech_conditioning_input):
|
|
"""
|
|
Randomly permute the conditioning spectrogram, to destroy any structure present. Note that since the
|
|
conditioning input is derived from a discrete spectrogram, it does actually retain structure, but only a little
|
|
bit (actually: exactly how much we want; enough to discriminate different vocal qualities, but nothing about
|
|
what is being said).
|
|
"""
|
|
cond_input = speech_conditioning_input[:,:,torch.randperm(speech_conditioning_input.shape[-1])]
|
|
if cond_input.shape[-1] > self.max_conditioning_length:
|
|
cond_input = cond_input[:,:,:self.max_conditioning_length]
|
|
return cond_input
|
|
|
|
|
|
def get_top_embeddings(self, embedded_input):
|
|
true_embeddings = self.top_encoder(embedded_input)
|
|
inputs = torch.cat([self.top_gpt_start_embedding, true_embeddings[:,:-1]], dim=1)
|
|
top_pred = self.top_gpt(inputs_embeds=inputs, return_dict=True)
|
|
return top_pred.last_hidden_state, true_embeddings
|
|
|
|
|
|
def inject_top_embeddings(self, embedded_input, probability_of_true_top_embedding=.5):
|
|
pred, true = self.get_top_embeddings(embedded_input)
|
|
rand = torch.bernoulli(torch.full((1,embedded_input.shape[1]),
|
|
fill_value=probability_of_true_top_embedding)).to(embedded_input.device)
|
|
mix = pred * rand + true * (not rand)
|
|
embs = torch.chunk(embedded_input, self.top_dim_reduction, dim=1)
|
|
assert len(embs) == mix.shape[1]
|
|
rejoin = []
|
|
for i, emb in enumerate(embs):
|
|
rejoin.append(torch.cat([mix[i], emb]), dim=1)
|
|
return torch.cat(rejoin, dim=1)
|
|
|
|
|
|
def get_logits(self, speech_conditioning_input, first_inputs, first_head, second_inputs=None, second_head=None, get_attns=False):
|
|
if second_inputs is not None:
|
|
emb = torch.cat([speech_conditioning_input, first_inputs, second_inputs], dim=1)
|
|
else:
|
|
emb = torch.cat([speech_conditioning_input, first_inputs], dim=1)
|
|
|
|
gpt_out = self.bottom_gpt(inputs_embeds=emb, return_dict=True, output_attentions=get_attns)
|
|
if get_attns:
|
|
return gpt_out.attentions
|
|
|
|
enc = gpt_out.last_hidden_state[:, 1:] # The first logit is tied to the speech_conditioning_input
|
|
enc = self.final_norm(enc)
|
|
first_logits = enc[:, :first_inputs.shape[1]]
|
|
first_logits = first_head(first_logits)
|
|
first_logits = first_logits.permute(0,2,1)
|
|
if second_inputs is not None:
|
|
second_logits = enc[:, -second_inputs.shape[1]:]
|
|
second_logits = second_head(second_logits)
|
|
second_logits = second_logits.permute(0,2,1)
|
|
return first_logits, second_logits
|
|
else:
|
|
return first_logits
|
|
|
|
def forward(self, speech_conditioning_input, text_inputs, mel_inputs, wav_lengths, text_first=True, return_attentions=False):
|
|
"""
|
|
Forward pass that uses both text and voice in either text conditioning mode or voice conditioning mode
|
|
(actuated by `text_first`).
|
|
|
|
speech_conditioning_input: MEL float tensor, (b,80,s)
|
|
text_inputs: long tensor, (b,t)
|
|
mel_inputs: long tensor, (b,m)
|
|
wav_lengths: long tensor, (b,)
|
|
"""
|
|
assert self.max_mel_tokens >= mel_inputs.shape[1], f'{mel_inputs.shape[1]}'
|
|
assert self.max_symbols_per_phrase >= text_inputs.shape[1], f'{text_inputs.shape[1]}'
|
|
assert self.max_total_tokens >= mel_inputs.shape[1] + text_inputs.shape[1], f'{mel_inputs.shape[1]}, {text_inputs.shape[1]}'
|
|
|
|
mel_inputs = self.set_mel_padding(mel_inputs, wav_lengths)
|
|
speech_conditioning_input = self.randomly_permute_conditioning_input(speech_conditioning_input)
|
|
speech_conditioning_input = self.conditioning_encoder(speech_conditioning_input).unsqueeze(1)
|
|
|
|
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_text_token, self.stop_text_token)
|
|
text_emb = self.text_embedding(text_inputs) + self.text_pos_paired_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
|
|
mel_inputs, mel_targets = self.build_aligned_inputs_and_targets(mel_inputs, self.start_mel_token, self.stop_mel_token)
|
|
mel_emb = self.bottom_gpt.get_input_embeddings()(mel_inputs)
|
|
mel_emb = mel_emb + self.mel_pos_paired_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
|
|
|
|
if text_first:
|
|
text_logits, mel_logits = self.get_logits(speech_conditioning_input, text_emb, self.text_head, mel_emb, self.mel_head, get_attns=return_attentions)
|
|
else:
|
|
mel_logits, text_logits = self.get_logits(speech_conditioning_input, mel_emb, self.mel_head, text_emb, self.text_head, get_attns=return_attentions)
|
|
|
|
if return_attentions:
|
|
return mel_logits
|
|
loss_text = F.cross_entropy(text_logits, text_targets.long())
|
|
loss_mel = F.cross_entropy(mel_logits, mel_targets.long())
|
|
return loss_text.mean(), loss_mel.mean(), mel_logits
|
|
|
|
def text_forward(self, speech_conditioning_input, text_inputs):
|
|
"""
|
|
Performs autoregressive modeling on only text. Still requires a speech_conditioning_input due to the way the
|
|
model inputs are formatted. Just provide any audio clip (arguably, zeros could be provided).
|
|
"""
|
|
assert self.max_symbols_per_phrase >= text_inputs.shape[1], f'{text_inputs.shape[1]}'
|
|
|
|
speech_conditioning_input = self.randomly_permute_conditioning_input(speech_conditioning_input)
|
|
speech_conditioning_input = self.conditioning_encoder(speech_conditioning_input).unsqueeze(1)
|
|
|
|
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_text_token, self.stop_text_token)
|
|
text_emb = self.text_embedding(text_inputs) + self.text_pos_solo_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
|
|
text_logits = self.get_logits(speech_conditioning_input, text_emb, self.text_head)
|
|
loss_text = F.cross_entropy(text_logits, text_targets.long())
|
|
return loss_text.mean()
|
|
|
|
def speech_forward(self, speech_conditioning_input, mel_inputs, wav_lengths):
|
|
"""
|
|
Performs autoregressive modeling on only speech data.
|
|
"""
|
|
assert self.max_mel_tokens >= mel_inputs.shape[1], f'{mel_inputs.shape[1]}'
|
|
|
|
mel_inputs = self.set_mel_padding(mel_inputs, wav_lengths)
|
|
speech_conditioning_input = self.randomly_permute_conditioning_input(speech_conditioning_input)
|
|
speech_conditioning_input = self.conditioning_encoder(speech_conditioning_input).unsqueeze(1)
|
|
|
|
mel_inputs, mel_targets = self.build_aligned_inputs_and_targets(mel_inputs, self.start_mel_token, self.stop_mel_token)
|
|
mel_emb = self.bottom_gpt.get_input_embeddings()(mel_inputs)
|
|
mel_emb = mel_emb + self.mel_pos_solo_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
|
|
mel_logits = self.get_logits(speech_conditioning_input, mel_emb, self.mel_head)
|
|
loss_mel = F.cross_entropy(mel_logits, mel_targets.long())
|
|
return loss_mel.mean()
|
|
|
|
def inference_speech(self, speech_conditioning_input, text_inputs, **hf_generate_kwargs):
|
|
if not hasattr(self, 'inference_model'):
|
|
self.inference_model = GPT2InferenceModel(self.bottom_gpt_config, self.bottom_gpt, self.mel_pos_paired_embedding, self.final_norm, self.mel_head)
|
|
|
|
text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_text_token, self.stop_text_token)
|
|
text_emb = self.text_embedding(text_inputs) + self.text_pos_paired_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
|
|
|
|
# Randomly permute the conditioning spectrogram, to destroy any structure present.
|
|
speech_conditioning_input = self.randomly_permute_conditioning_input(speech_conditioning_input)
|
|
cond = self.conditioning_encoder(speech_conditioning_input).unsqueeze(1)
|
|
|
|
emb = torch.cat([cond, text_emb], dim=1)
|
|
self.inference_model.store_mel_emb(emb)
|
|
|
|
fake_inputs = torch.full((emb.shape[0],emb.shape[1]+1,), fill_value=1, dtype=torch.long, device=text_inputs.device)
|
|
fake_inputs[:,-1] = self.start_mel_token
|
|
|
|
gen = self.inference_model.generate(fake_inputs, bos_token_id=self.start_mel_token, pad_token_id=self.stop_mel_token, eos_token_id=self.stop_mel_token,
|
|
max_length=self.bottom_gpt_config.n_positions, **hf_generate_kwargs)
|
|
return gen[:, fake_inputs.shape[1]:]
|
|
|
|
|
|
@register_model
|
|
def register_unified_gpt_voice_bilevel(opt_net, opt):
|
|
return UnifiedGptVoice(**opt_get(opt_net, ['kwargs'], {}))
|
|
|
|
|
|
if __name__ == '__main__':
|
|
gpt = UnifiedGptVoice(model_dim=256, heads=4)
|
|
l = gpt(torch.randn(2, 80, 800),
|
|
torch.randint(high=len(symbols), size=(2,80)),
|
|
torch.randint(high=8192, size=(2,250)),
|
|
torch.tensor([150*256,195*256]))
|