378 lines
22 KiB
Python
378 lines
22 KiB
Python
import os
|
|
import os.path as osp
|
|
import random
|
|
|
|
import torch
|
|
import torchaudio
|
|
import torchvision.utils
|
|
from pytorch_fid.fid_score import calculate_frechet_distance
|
|
from torch import distributed
|
|
from tqdm import tqdm
|
|
from transformers import Wav2Vec2ForCTC
|
|
import torch.nn.functional as F
|
|
import numpy as np
|
|
|
|
import trainer.eval.evaluator as evaluator
|
|
from data.audio.paired_voice_audio_dataset import load_tsv_aligned_codes
|
|
from data.audio.unsupervised_audio_dataset import load_audio
|
|
from data.audio.voice_tokenizer import VoiceBpeTokenizer
|
|
from models.clip.mel_text_clip import MelTextCLIP
|
|
from models.audio.tts.tacotron2 import text_to_sequence
|
|
from scripts.audio.gen.speech_synthesis_utils import load_discrete_vocoder_diffuser, wav_to_mel, load_speech_dvae, \
|
|
convert_mel_to_codes, load_univnet_vocoder, wav_to_univnet_mel, load_clvp
|
|
from trainer.injectors.audio_injectors import denormalize_mel, TorchMelSpectrogramInjector, normalize_mel
|
|
from utils.util import ceil_multiple, opt_get, load_model_from_config, pad_or_truncate
|
|
|
|
|
|
class AudioDiffusionFid(evaluator.Evaluator):
|
|
"""
|
|
Evaluator produces generate from a diffusion model, then uses a CLIP model to judge the similarity between text & speech.
|
|
|
|
This evaluator is kind of a mess. It has been repeatedly modified to work with several different model types, which
|
|
means it is bloated beyond belief. I would not recommend attempting to understand what is going on here.
|
|
"""
|
|
def __init__(self, model, opt_eval, env):
|
|
super().__init__(model, opt_eval, env, uses_all_ddp=True)
|
|
self.real_path = opt_eval['eval_tsv']
|
|
self.data = load_tsv_aligned_codes(self.real_path)
|
|
|
|
# Deterministically shuffle the data.
|
|
ostate = random.getstate()
|
|
random.seed(5)
|
|
random.shuffle(self.data)
|
|
random.setstate(ostate)
|
|
|
|
if 'clip_dataset' in opt_eval.keys():
|
|
self.data = self.data[:opt_eval['clip_dataset']]
|
|
if distributed.is_initialized() and distributed.get_world_size() > 1:
|
|
self.skip = distributed.get_world_size() # One batch element per GPU.
|
|
else:
|
|
self.skip = 1
|
|
diffusion_steps = opt_get(opt_eval, ['diffusion_steps'], 50)
|
|
diffusion_schedule = opt_get(env['opt'], ['steps', 'generator', 'injectors', 'diffusion', 'beta_schedule', 'schedule_name'], None)
|
|
if diffusion_schedule is None:
|
|
print("Unable to infer diffusion schedule from master options. Getting it from eval (or guessing).")
|
|
diffusion_schedule = opt_get(opt_eval, ['diffusion_schedule'], 'cosine')
|
|
conditioning_free_diffusion_enabled = opt_get(opt_eval, ['conditioning_free'], False)
|
|
conditioning_free_k = opt_get(opt_eval, ['conditioning_free_k'], 1)
|
|
self.diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=diffusion_steps, schedule=diffusion_schedule,
|
|
enable_conditioning_free_guidance=conditioning_free_diffusion_enabled,
|
|
conditioning_free_k=conditioning_free_k)
|
|
self.bpe_tokenizer = VoiceBpeTokenizer('../experiments/bpe_lowercase_asr_256.json')
|
|
self.dev = self.env['device']
|
|
mode = opt_get(opt_eval, ['diffusion_type'], 'tts')
|
|
self.local_modules = {}
|
|
if mode == 'tts':
|
|
self.diffusion_fn = self.perform_diffusion_tts
|
|
elif mode == 'original_vocoder':
|
|
self.local_modules['dvae'] = load_speech_dvae().cpu()
|
|
self.diffusion_fn = self.perform_original_diffusion_vocoder
|
|
elif mode == 'vocoder':
|
|
self.local_modules['dvae'] = load_speech_dvae().cpu()
|
|
self.diffusion_fn = self.perform_diffusion_vocoder
|
|
elif mode == 'ctc_to_mel':
|
|
self.diffusion_fn = self.perform_diffusion_ctc
|
|
self.local_modules['vocoder'] = load_univnet_vocoder().cpu()
|
|
self.local_modules['clvp'] = load_clvp()
|
|
elif 'tts9_mel' in mode:
|
|
mel_means, self.mel_max, self.mel_min, mel_stds, mel_vars = torch.load('../experiments/univnet_mel_norms.pth')
|
|
self.local_modules['dvae'] = load_speech_dvae().cpu()
|
|
self.local_modules['vocoder'] = load_univnet_vocoder().cpu()
|
|
self.diffusion_fn = self.perform_diffusion_tts9_mel_from_codes
|
|
if mode == 'tts9_mel_autoin':
|
|
self.local_modules['autoregressive'] = load_model_from_config("../experiments/train_gpt_tts_unified.yml",
|
|
model_name='gpt',
|
|
also_load_savepoint=False,
|
|
load_path='../experiments/unified_large_diverse_basis.pth',
|
|
device=torch.device('cpu')).cuda().eval()
|
|
self.tts9_codegen = self.tts9_get_autoregressive_codes
|
|
else:
|
|
self.tts9_codegen = self.tts9_get_dvae_codes
|
|
elif 'tfd' == mode:
|
|
self.diffusion_fn = self.perform_diffusion_tfd
|
|
self.local_modules['vocoder'] = load_univnet_vocoder().cpu()
|
|
elif 'tfd_ar' == mode:
|
|
self.local_modules['dvae'] = load_speech_dvae().cpu()
|
|
self.local_modules['autoregressive'] = load_model_from_config("../experiments/train_gpt_tts_unified.yml",
|
|
model_name='gpt',
|
|
also_load_savepoint=False,
|
|
load_path='../experiments/tortoise_ar.pth',
|
|
device=torch.device('cpu')).cuda().eval()
|
|
self.diffusion_fn = self.perform_diffusion_tfd_ar_prior
|
|
self.local_modules['vocoder'] = load_univnet_vocoder().cpu()
|
|
|
|
def perform_diffusion_tts(self, audio, codes, text, sample_rate=5500):
|
|
real_resampled = torchaudio.functional.resample(audio, 22050, sample_rate).unsqueeze(0)
|
|
aligned_codes_compression_factor = sample_rate * 221 // 11025
|
|
output_size = codes.shape[-1]*aligned_codes_compression_factor
|
|
padded_size = ceil_multiple(output_size, 2048)
|
|
padding_added = padded_size - output_size
|
|
padding_needed_for_codes = padding_added // aligned_codes_compression_factor
|
|
if padding_needed_for_codes > 0:
|
|
codes = F.pad(codes, (0, padding_needed_for_codes))
|
|
output_shape = (1, 1, padded_size)
|
|
gen = self.diffuser.p_sample_loop(self.model, output_shape,
|
|
model_kwargs={'tokens': codes.unsqueeze(0),
|
|
'conditioning_input': real_resampled})
|
|
return gen, real_resampled, sample_rate
|
|
|
|
def perform_original_diffusion_vocoder(self, audio, codes, text, sample_rate=11025):
|
|
mel = wav_to_mel(audio)
|
|
mel_codes = convert_mel_to_codes(self.local_modules['dvae'], mel)
|
|
back_to_mel = self.local_modules['dvae'].decode(mel_codes)[0]
|
|
orig_audio = audio
|
|
real_resampled = torchaudio.functional.resample(audio, 22050, sample_rate).unsqueeze(0)
|
|
|
|
output_size = real_resampled.shape[-1]
|
|
aligned_mel_compression_factor = output_size // back_to_mel.shape[-1]
|
|
padded_size = ceil_multiple(output_size, 2048)
|
|
padding_added = padded_size - output_size
|
|
padding_needed_for_codes = padding_added // aligned_mel_compression_factor
|
|
if padding_needed_for_codes > 0:
|
|
back_to_mel = F.pad(back_to_mel, (0, padding_needed_for_codes))
|
|
output_shape = (1, 1, padded_size)
|
|
gen = self.diffuser.p_sample_loop(self.model, output_shape,
|
|
model_kwargs={'spectrogram': back_to_mel,
|
|
'conditioning_input': orig_audio.unsqueeze(0)})
|
|
|
|
# Pop it back down to 5.5kHz for an accurate comparison with the other diffusers.
|
|
real_resampled = torchaudio.functional.resample(real_resampled.squeeze(0), sample_rate, 5500).unsqueeze(0)
|
|
gen = torchaudio.functional.resample(gen.squeeze(0), sample_rate, 5500).unsqueeze(0)
|
|
return gen, real_resampled, 5500
|
|
|
|
|
|
def perform_diffusion_vocoder(self, audio, codes, text, sample_rate=5500):
|
|
mel = wav_to_mel(audio)
|
|
mel_codes = convert_mel_to_codes(self.local_modules['dvae'], mel)
|
|
text_codes = text_to_sequence(text)
|
|
real_resampled = torchaudio.functional.resample(audio, 22050, sample_rate).unsqueeze(0)
|
|
|
|
output_size = real_resampled.shape[-1]
|
|
aligned_codes_compression_factor = output_size // mel_codes.shape[-1]
|
|
padded_size = ceil_multiple(output_size, 2048)
|
|
padding_added = padded_size - output_size
|
|
padding_needed_for_codes = padding_added // aligned_codes_compression_factor
|
|
if padding_needed_for_codes > 0:
|
|
mel_codes = F.pad(mel_codes, (0, padding_needed_for_codes))
|
|
output_shape = (1, 1, padded_size)
|
|
gen = self.diffuser.p_sample_loop(self.model, output_shape,
|
|
model_kwargs={'tokens': mel_codes,
|
|
'conditioning_input': audio.unsqueeze(0),
|
|
'unaligned_input': torch.tensor(text_codes, device=audio.device).unsqueeze(0)})
|
|
return gen, real_resampled, sample_rate
|
|
|
|
def tts9_get_autoregressive_codes(self, mel, text):
|
|
mel_codes = convert_mel_to_codes(self.local_modules['dvae'], mel)
|
|
text_codes = torch.LongTensor(self.bpe_tokenizer.encode(text)).unsqueeze(0).to(mel.device)
|
|
cond_inputs = pad_or_truncate(mel, 132300//256).unsqueeze(1)
|
|
mlc = self.local_modules['autoregressive'].mel_length_compression
|
|
auto_latents = self.local_modules['autoregressive'](cond_inputs, text_codes,
|
|
torch.tensor([text_codes.shape[-1]], device=mel.device),
|
|
mel_codes,
|
|
torch.tensor([mel_codes.shape[-1]*mlc], device=mel.device),
|
|
text_first=True, raw_mels=None, return_latent=True)
|
|
return auto_latents
|
|
|
|
def tts9_get_dvae_codes(self, mel, text):
|
|
return convert_mel_to_codes(self.local_modules['dvae'], mel)
|
|
|
|
def perform_diffusion_tts9_mel_from_codes(self, audio, codes, text):
|
|
SAMPLE_RATE = 24000
|
|
mel = wav_to_mel(audio)
|
|
mel_codes = self.tts9_codegen(mel, text)
|
|
real_resampled = torchaudio.functional.resample(audio, 22050, SAMPLE_RATE).unsqueeze(0)
|
|
univnet_mel = wav_to_univnet_mel(real_resampled, do_normalization=False) # to be used for a conditioning input, but also guides output shape.
|
|
output_shape = univnet_mel.shape
|
|
gen_mel = self.diffuser.p_sample_loop(self.model, output_shape,
|
|
model_kwargs={'aligned_conditioning': mel_codes,
|
|
'conditioning_input': univnet_mel})
|
|
# denormalize mel
|
|
gen_mel = denormalize_mel(gen_mel)
|
|
|
|
gen_wav = self.local_modules['vocoder'].inference(gen_mel)
|
|
real_dec = self.local_modules['vocoder'].inference(univnet_mel)
|
|
return gen_wav.float(), real_dec, gen_mel, univnet_mel, SAMPLE_RATE
|
|
|
|
def perform_diffusion_ctc(self, audio, codes, text):
|
|
SAMPLE_RATE = 24000
|
|
text_codes = torch.LongTensor(self.bpe_tokenizer.encode(text)).unsqueeze(0).to(audio.device)
|
|
clvp_latent = self.local_modules['clvp'].embed_text(text_codes)
|
|
|
|
real_resampled = torchaudio.functional.resample(audio, 22050, SAMPLE_RATE).unsqueeze(0)
|
|
univnet_mel = wav_to_univnet_mel(real_resampled, do_normalization=True)
|
|
output_shape = univnet_mel.shape
|
|
cond_mel = TorchMelSpectrogramInjector({'n_mel_channels': 100, 'mel_fmax': 11000, 'filter_length': 8000, 'normalize': True,
|
|
'true_normalization': True, 'in': 'in', 'out': 'out'}, {})({'in': audio})['out']
|
|
|
|
gen_mel = self.diffuser.p_sample_loop(self.model, output_shape, model_kwargs={'codes': codes.unsqueeze(0),
|
|
'conditioning_input': cond_mel, 'type': torch.tensor([0], device=codes.device),
|
|
'clvp_input': clvp_latent})
|
|
gen_mel_denorm = denormalize_mel(gen_mel)
|
|
|
|
gen_wav = self.local_modules['vocoder'].inference(gen_mel_denorm)
|
|
real_dec = self.local_modules['vocoder'].inference(denormalize_mel(univnet_mel))
|
|
return gen_wav.float(), real_dec, gen_mel, univnet_mel, SAMPLE_RATE
|
|
|
|
def perform_diffusion_tfd(self, audio, codes, text):
|
|
SAMPLE_RATE = 24000
|
|
audio_resampled = torchaudio.functional.resample(audio, 22050, SAMPLE_RATE).unsqueeze(0)
|
|
vmel = wav_to_mel(audio)
|
|
umel = wav_to_univnet_mel(audio_resampled, do_normalization=True)
|
|
gen_mel = self.diffuser.p_sample_loop(self.model, umel.shape,
|
|
model_kwargs={'truth_mel': vmel,
|
|
'conditioning_input': None,
|
|
'disable_diversity': True})
|
|
|
|
gen_wav = self.local_modules['vocoder'].inference(denormalize_mel(gen_mel))
|
|
real_dec = self.local_modules['vocoder'].inference(denormalize_mel(umel))
|
|
return gen_wav.float(), real_dec, gen_mel, umel, SAMPLE_RATE
|
|
|
|
def perform_diffusion_tfd_ar_prior(self, audio, codes, text):
|
|
SAMPLE_RATE = 24000
|
|
audio_resampled = torchaudio.functional.resample(audio, 22050, SAMPLE_RATE).unsqueeze(0)
|
|
vmel = wav_to_mel(audio)
|
|
umel = wav_to_univnet_mel(audio_resampled, do_normalization=True)
|
|
|
|
mel_codes = convert_mel_to_codes(self.local_modules['dvae'], vmel)
|
|
text_codes = torch.LongTensor(self.bpe_tokenizer.encode(text)).unsqueeze(0).to(vmel.device)
|
|
cond_inputs = pad_or_truncate(vmel, 132300//256).unsqueeze(1)
|
|
mlc = self.local_modules['autoregressive'].mel_length_compression
|
|
auto_latents = self.local_modules['autoregressive'](cond_inputs, text_codes,
|
|
torch.tensor([text_codes.shape[-1]], device=vmel.device),
|
|
mel_codes,
|
|
torch.tensor([mel_codes.shape[-1]*mlc], device=vmel.device),
|
|
text_first=True, raw_mels=None, return_latent=True)
|
|
|
|
gen_mel = self.diffuser.p_sample_loop(self.model, umel.shape,
|
|
model_kwargs={'codes': auto_latents})
|
|
|
|
gen_wav = self.local_modules['vocoder'].inference(denormalize_mel(gen_mel))
|
|
real_dec = self.local_modules['vocoder'].inference(denormalize_mel(umel))
|
|
return gen_wav.float(), real_dec, gen_mel, umel, SAMPLE_RATE
|
|
|
|
def load_projector(self):
|
|
"""
|
|
Builds the CLIP model used to project speech into a latent. This model has fixed parameters and a fixed loading
|
|
path for the time being.
|
|
"""
|
|
model = MelTextCLIP(dim_text=512, dim_latent=512, dim_speech=512, num_text_tokens=148, text_enc_depth=8,
|
|
text_seq_len=400, text_heads=8, speech_enc_depth=10, speech_heads=8, speech_seq_len=1000,
|
|
text_mask_percentage=.15, voice_mask_percentage=.15)
|
|
weights = torch.load('../experiments/clip_text_to_voice_for_speech_fid.pth')
|
|
model.load_state_dict(weights)
|
|
return model
|
|
|
|
def project(self, projector, sample, sample_rate):
|
|
sample = torchaudio.functional.resample(sample, sample_rate, 22050)
|
|
mel = wav_to_mel(sample)
|
|
return projector.get_speech_projection(mel).squeeze(0) # Getting rid of the batch dimension means it's just [hidden_dim]
|
|
|
|
def load_w2v(self):
|
|
return Wav2Vec2ForCTC.from_pretrained("jbetker/wav2vec2-large-robust-ft-libritts-voxpopuli")
|
|
|
|
def intelligibility_loss(self, w2v, sample, real_sample, sample_rate, real_text):
|
|
"""
|
|
Measures the differences between CTC losses using wav2vec2 against the real sample and the generated sample.
|
|
"""
|
|
text_codes = torch.tensor(text_to_sequence(real_text), device=sample.device)
|
|
results = []
|
|
for s in [sample, real_sample]:
|
|
s = torchaudio.functional.resample(s, sample_rate, 16000)
|
|
norm_s = (s - s.mean()) / torch.sqrt(s.var() + 1e-7)
|
|
norm_s = norm_s.squeeze(1)
|
|
loss = w2v(input_values=norm_s, labels=text_codes).loss
|
|
results.append(loss)
|
|
gen_loss, real_loss = results
|
|
return gen_loss - real_loss
|
|
|
|
def compute_frechet_distance(self, proj1, proj2):
|
|
# I really REALLY FUCKING HATE that this is going to numpy. Why does "pytorch_fid" operate in numpy land. WHY?
|
|
proj1 = proj1.cpu().numpy()
|
|
proj2 = proj2.cpu().numpy()
|
|
mu1 = np.mean(proj1, axis=0)
|
|
mu2 = np.mean(proj2, axis=0)
|
|
sigma1 = np.cov(proj1, rowvar=False)
|
|
sigma2 = np.cov(proj2, rowvar=False)
|
|
return torch.tensor(calculate_frechet_distance(mu1, sigma1, mu2, sigma2))
|
|
|
|
def perform_eval(self):
|
|
save_path = osp.join(self.env['base_path'], "../", "audio_eval", str(self.env["step"]))
|
|
os.makedirs(save_path, exist_ok=True)
|
|
|
|
projector = self.load_projector().to(self.env['device'])
|
|
projector.eval()
|
|
|
|
w2v = self.load_w2v().to(self.env['device'])
|
|
w2v.eval()
|
|
for k, mod in self.local_modules.items():
|
|
self.local_modules[k] = mod.to(self.env['device'])
|
|
|
|
# Attempt to fix the random state as much as possible. RNG state will be restored before returning.
|
|
rng_state = torch.get_rng_state()
|
|
torch.manual_seed(5)
|
|
self.model.eval()
|
|
|
|
with torch.no_grad():
|
|
gen_projections = []
|
|
real_projections = []
|
|
intelligibility_losses = []
|
|
for i in tqdm(list(range(0, len(self.data), self.skip))):
|
|
path, text, codes = self.data[(i + self.env['rank']) % len(self.data)]
|
|
audio = load_audio(path, 22050).to(self.dev)
|
|
codes = codes.to(self.dev)
|
|
sample, ref, gen_mel, ref_mel, sample_rate = self.diffusion_fn(audio, codes, text)
|
|
|
|
gen_projections.append(self.project(projector, sample, sample_rate).cpu()) # Store on CPU to avoid wasting GPU memory.
|
|
real_projections.append(self.project(projector, ref, sample_rate).cpu())
|
|
intelligibility_losses.append(self.intelligibility_loss(w2v, sample, ref, sample_rate, text))
|
|
|
|
torchvision.utils.save_image((gen_mel.unsqueeze(1) + 1) / 2, os.path.join(save_path, f'{self.env["rank"]}_{i}_mel.png'))
|
|
torchvision.utils.save_image((ref_mel.unsqueeze(1) + 1) / 2, os.path.join(save_path, f'{self.env["rank"]}_{i}_mel_target.png'))
|
|
torchaudio.save(os.path.join(save_path, f"{self.env['rank']}_{i}_gen.wav"), sample.squeeze(0).cpu(), sample_rate)
|
|
torchaudio.save(os.path.join(save_path, f"{self.env['rank']}_{i}_real.wav"), ref.squeeze(0).cpu(), sample_rate)
|
|
gen_projections = torch.stack(gen_projections, dim=0)
|
|
real_projections = torch.stack(real_projections, dim=0)
|
|
intelligibility_loss = torch.stack(intelligibility_losses, dim=0).mean()
|
|
frechet_distance = torch.tensor(self.compute_frechet_distance(gen_projections, real_projections), device=self.env['device'])
|
|
|
|
if distributed.is_initialized() and distributed.get_world_size() > 1:
|
|
distributed.all_reduce(frechet_distance)
|
|
frechet_distance = frechet_distance / distributed.get_world_size()
|
|
distributed.all_reduce(intelligibility_loss)
|
|
intelligibility_loss = intelligibility_loss / distributed.get_world_size()
|
|
|
|
self.model.train()
|
|
torch.set_rng_state(rng_state)
|
|
|
|
# Put modules used for evaluation back into CPU memory.
|
|
for k, mod in self.local_modules.items():
|
|
self.local_modules[k] = mod.cpu()
|
|
|
|
return {"frechet_distance": frechet_distance, "intelligibility_loss": intelligibility_loss}
|
|
|
|
"""
|
|
if __name__ == '__main__':
|
|
from utils.util import load_model_from_config
|
|
|
|
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_diffusion_tts7_dvae_thin_with_text.yml', 'generator',
|
|
also_load_savepoint=False, load_path='X:\\dlas\\experiments\\train_diffusion_tts7_dvae_thin_with_text\\models\\56500_generator_ema.pth').cuda()
|
|
opt_eval = {'eval_tsv': 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv', 'diffusion_steps': 100,
|
|
'conditioning_free': False, 'conditioning_free_k': 1,
|
|
'diffusion_schedule': 'linear', 'diffusion_type': 'vocoder'}
|
|
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval', 'step': 2, 'device': 'cuda', 'opt': {}}
|
|
eval = AudioDiffusionFid(diffusion, opt_eval, env)
|
|
print(eval.perform_eval())
|
|
"""
|
|
|
|
|
|
if __name__ == '__main__':
|
|
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_tts_diffusion_tfd12_ar_inputs.yml', 'generator',
|
|
also_load_savepoint=False,
|
|
load_path='X:\\dlas\\experiments\\train_tts_diffusion_tfd12_ar_inputs_pretrain\\models\\4500_generator.pth').cuda()
|
|
opt_eval = {'eval_tsv': 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv', 'diffusion_steps': 50,
|
|
'conditioning_free': False, 'conditioning_free_k': 1,
|
|
'diffusion_schedule': 'linear', 'diffusion_type': 'tfd_ar'}
|
|
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval', 'step': 101, 'device': 'cuda', 'opt': {}}
|
|
eval = AudioDiffusionFid(diffusion, opt_eval, env)
|
|
print(eval.perform_eval())
|