DL-Art-School/codes/models/gpt_voice/gpt_tts.py

178 lines
8.6 KiB
Python
Raw Normal View History

2021-07-28 02:33:30 +00:00
import torch
import torch.nn as nn
import torch.nn.functional as F
2021-08-09 17:56:06 +00:00
from munch import munchify
2021-07-28 02:33:30 +00:00
2021-08-05 02:07:45 +00:00
from models.gpt_voice.lucidrains_gpt import Transformer
from models.gpt_voice.min_gpt import GPT, GPTConfig
2021-07-28 02:33:30 +00:00
from models.tacotron2.taco_utils import get_mask_from_lengths
from models.tacotron2.text import symbols
from trainer.networks import register_model
2021-08-07 04:08:51 +00:00
from utils.util import opt_get
2021-07-28 02:33:30 +00:00
class GptTts(nn.Module):
2021-08-05 02:07:45 +00:00
MAX_SYMBOLS_PER_PHRASE = 200
NUMBER_SYMBOLS = len(symbols)
NUMBER_TEXT_TOKENS = NUMBER_SYMBOLS + MAX_SYMBOLS_PER_PHRASE + 2
2021-08-07 04:01:06 +00:00
MEL_DICTIONARY_SIZE = 512+3
2021-08-04 03:08:27 +00:00
MEL_START_TOKEN = MEL_DICTIONARY_SIZE-3
MEL_STOP_TOKEN = MEL_DICTIONARY_SIZE-2
2021-08-07 04:08:51 +00:00
def __init__(self, layers=8, model_dim=512, heads=8):
2021-07-28 02:33:30 +00:00
super().__init__()
2021-08-07 04:01:06 +00:00
max_mel_frames = 900 * 1 // 4 # 900 is the max number of MEL frames. The VQVAE outputs 1/8 of the input mel as tokens.
2021-07-28 02:33:30 +00:00
2021-07-31 05:07:35 +00:00
self.model_dim = model_dim
self.max_mel_frames = max_mel_frames
2021-08-05 02:07:45 +00:00
self.text_embedding = nn.Embedding(self.NUMBER_TEXT_TOKENS, model_dim)
2021-08-04 03:08:27 +00:00
self.mel_embedding = nn.Embedding(self.MEL_DICTIONARY_SIZE, model_dim)
2021-08-05 02:07:45 +00:00
self.text_pos_embedding = nn.Embedding(self.MAX_SYMBOLS_PER_PHRASE, model_dim)
2021-08-04 03:08:27 +00:00
self.mel_pos_embedding = nn.Embedding(max_mel_frames, model_dim)
self.gpt = Transformer(dim=model_dim, depth=layers, seq_len=1+self.MAX_SYMBOLS_PER_PHRASE+max_mel_frames, heads=heads,
attn_dropout=.1, ff_dropout=.1)
2021-08-04 03:08:27 +00:00
self.final_norm = nn.LayerNorm(model_dim)
2021-08-05 02:07:45 +00:00
self.text_head = nn.Linear(model_dim, self.NUMBER_TEXT_TOKENS)
2021-08-04 03:08:27 +00:00
self.mel_head = nn.Linear(model_dim, self.MEL_DICTIONARY_SIZE)
def forward(self, text_inputs, text_lengths, mel_targets, output_lengths):
2021-07-28 02:33:30 +00:00
text_emb = self.text_embedding(text_inputs)
2021-08-04 03:08:27 +00:00
text_emb = text_emb + self.text_pos_embedding(torch.arange(text_inputs.shape[1], device=text_inputs.device))
mel_emb = self.mel_embedding(mel_targets)
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_targets.shape[1], device=mel_targets.device))
emb = torch.cat([text_emb, mel_emb], dim=1)
enc = self.gpt(emb)
# Compute logits for text and mel heads
text_logits = self.final_norm(enc[:, :text_emb.shape[1]])
mel_logits = self.final_norm(enc[:, text_emb.shape[1]:])
2021-08-05 02:07:45 +00:00
text_logits = self.text_head(text_logits)
2021-08-04 03:08:27 +00:00
mel_logits = self.mel_head(mel_logits)
# Compute loss
text_targets = text_inputs[:,1:]
text_logits = text_logits.permute(0,2,1)[:,:,:-1] # The last element of the logits is unneeded because the input to the transformer contains a <EOS> token for both text and mel.
loss_text = F.cross_entropy(text_logits, text_targets, reduction='none')
mel_targets = mel_targets[:,1:]
mel_logits = mel_logits.permute(0,2,1)[:,:,:-1]
loss_mel = F.cross_entropy(mel_logits, mel_targets, reduction='none')
2021-08-04 03:08:27 +00:00
# Fix up mel_logits so it can go into a VAE decoder as well.
mel_codes = torch.argmax(F.softmax(mel_logits, dim=1), dim=1)
2021-08-05 02:07:45 +00:00
mel_pad_mask = ~get_mask_from_lengths(output_lengths-1, mel_targets.shape[1])
mel_codes = mel_codes * torch.ones_like(mel_codes).masked_fill_(mel_pad_mask, 0)
2021-08-05 02:07:45 +00:00
mel_codes = mel_codes[:,:-1] # Strip off <EOS> token too (or padding). The important part is that the output sequence length is identical to the VAE input.
2021-08-04 03:08:27 +00:00
extra_mask = mel_codes < self.MEL_DICTIONARY_SIZE-3 # The VAE doesn't know about START/STOP/PAD
mel_codes = mel_codes * extra_mask
2021-08-05 02:07:45 +00:00
# This class also returns the mel_targets for validation purposes. Format those.
mel_targets = mel_targets[:,:-1]
mel_targets = mel_targets * (mel_targets < self.MEL_DICTIONARY_SIZE-3)
return loss_text.mean(), loss_mel.mean(), mel_codes, mel_targets
2021-07-31 05:07:35 +00:00
def inference(self, text_inputs):
b, s = text_inputs.shape
2021-07-31 05:07:35 +00:00
text_emb = self.text_embedding(text_inputs)
text_emb = text_emb + self.text_pos_embedding(torch.arange(s, device=text_inputs.device))
mel_seq = torch.full((b,1), fill_value=self.MEL_START_TOKEN, device=text_emb.device)
stop_encountered = torch.zeros((b,), device=text_emb.device)
2021-08-05 12:46:30 +00:00
while not torch.all(stop_encountered) and len(mel_seq) < self.max_mel_frames:
mel_emb = self.mel_embedding(mel_seq)
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
emb = torch.cat([text_emb, mel_emb], dim=1)
enc = self.gpt(emb)
mel_logits = self.final_norm(enc[:, text_emb.shape[1]:])
mel_logits = self.mel_head(mel_logits)
mel_codes = torch.argmax(F.softmax(mel_logits, dim=-1), dim=-1)
2021-08-05 12:46:30 +00:00
mel_seq = torch.cat([mel_seq, mel_codes[:, -1].unsqueeze(1)], dim=1)
stop_encountered = torch.logical_or(stop_encountered, mel_seq[:,-1] == self.MEL_STOP_TOKEN)
if len(mel_seq) >= self.max_mel_frames:
print("Warning! Encountered frame limit before a stop token. Output is likely wrong.")
# Format mel_seq so that the DVAE can actually use it (it is a two-tiered DVAE)
mel_seq = mel_seq[:, 1:-1] # Remove first and last tokens, which were artificially added for GPT
mel_seq = mel_seq * (mel_seq < 512) # The DVAE doesn't understand BOS/EOS/PAD tokens.
return mel_seq
2021-07-31 05:07:35 +00:00
2021-08-09 17:56:06 +00:00
def inference_beam_topk(self, text):
def topk_sampler(distribution, k):
return torch.topk(distribution, k=k, dim=-1)
return self.inference_beam(text, topk_sampler)
def inference_beam_sampled(self, text):
def multinomial_sampler(distribution, k):
indices = torch.multinomial(distribution, num_samples=k, replacement=False)
values = torch.gather(distribution, dim=1, index=indices)
class container:
def __init__(self, i, v):
self.indices = i
self.values = v
return container(indices, values)
return self.inference_beam(text, multinomial_sampler)
def inference_beam(self, text_inputs, sampler_fn):
beam_width = 16
2021-08-09 17:56:06 +00:00
temperature = .8
b, s = text_inputs.shape
assert b == 1 # Beam search only works on batches of one.
text_emb = self.text_embedding(text_inputs)
text_emb = text_emb + self.text_pos_embedding(torch.arange(s, device=text_inputs.device))
mel_seq = torch.full((b,1), fill_value=self.MEL_START_TOKEN, device=text_emb.device)
probabilities = torch.ones((b,), device=text_emb.device)
while len(mel_seq) < self.max_mel_frames:
mel_emb = self.mel_embedding(mel_seq)
mel_emb = mel_emb + self.mel_pos_embedding(torch.arange(mel_emb.shape[1], device=mel_emb.device))
if text_emb.shape[0] != mel_emb.shape[0]:
text_emb = text_emb.repeat(mel_emb.shape[0], 1, 1)
emb = torch.cat([text_emb, mel_emb], dim=1)
enc = self.gpt(emb)
mel_logits = self.final_norm(enc[:, text_emb.shape[1]:])
mel_logits = self.mel_head(mel_logits)
2021-08-09 17:56:06 +00:00
topk = sampler_fn(F.softmax(temperature * mel_logits[:, -1], dim=-1), k=beam_width)
probabilities = (probabilities.repeat_interleave(beam_width, dim=0) * topk.values.flatten())
probabilities, sort_indices = torch.sort(probabilities, descending=True)
probabilities = probabilities[:beam_width]
mel_seq = mel_seq.repeat_interleave(beam_width, dim=0)
codes = topk.indices.flatten()
mel_seq = torch.cat([mel_seq, codes.unsqueeze(1)], dim=1)
mel_seq = mel_seq[sort_indices]
mel_seq = mel_seq[:beam_width]
if torch.all(torch.any(mel_seq == self.MEL_STOP_TOKEN, dim=1)):
break
if mel_seq.shape[1] >= self.max_mel_frames:
print("Warning! Encountered frame limit before a stop token. Output is likely wrong.")
# Format mel_seq so that the DVAE can actually use it (it is a two-tiered DVAE)
mel_seq = mel_seq[0, 1:-1].unsqueeze(0) # Pick most likely outcome, remove first and last tokens, which were artificially added for GPT
mel_seq = mel_seq * (mel_seq < 512) # The DVAE doesn't understand BOS/EOS/PAD tokens.
return mel_seq
2021-07-28 02:33:30 +00:00
@register_model
def register_gpt_tts(opt_net, opt):
2021-08-07 04:10:18 +00:00
return GptTts(**opt_get(opt_net, ['kwargs'], {}))
2021-07-28 02:33:30 +00:00
if __name__ == '__main__':
gpt = GptTts()
2021-08-04 03:08:27 +00:00
l1, l2, i = gpt(torch.randint(high=24, size=(2,60)),
torch.tensor([55,58]),
torch.randint(high=512, size=(2,310)),
torch.tensor([300,305]))
print(i.shape)
2021-07-31 05:07:35 +00:00
2021-07-31 21:57:57 +00:00
#o = gpt.infer(torch.randint(high=24, size=(2,60)))
#print(o.shape)
2021-07-31 05:07:35 +00:00