deviandice/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

MDF

Browse Source 
around and around in circles........
This commit is contained in:
James Betker 2022-05-01 23:04:56 -06:00
parent b712d3b72b
commit f4254609c1
1 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
codes/trainer/eval/music_diffusion_fid.py
View File

@ -4,6 +4,7 @@ from glob import glob
import torch import torch
import torchaudio import torchaudio
import torchvision
from pytorch_fid.fid_score import calculate_frechet_distance from pytorch_fid.fid_score import calculate_frechet_distance
from torch import distributed from torch import distributed
from tqdm import tqdm from tqdm import tqdm
@ -52,7 +53,7 @@ class MusicDiffusionFid(evaluator.Evaluator):
self.local_modules = {} self.local_modules = {}
if mode == 'standard': if mode == 'standard':
self.diffusion_fn = self.perform_diffusion_standard self.diffusion_fn = self.perform_diffusion_standard
self.spec_fn = TorchMelSpectrogramInjector({'n_mel_channels': 128, 'mel_fmax': 22000, 'normalize': True, 'in': 'in', 'out': 'out'}, {}) self.spec_fn = TorchMelSpectrogramInjector({'n_mel_channels': 256, 'mel_fmax': 22000, 'normalize': True, 'in': 'in', 'out': 'out'}, {})
def load_data(self, path): def load_data(self, path):
return list(glob(f'{path}/*.wav')) return list(glob(f'{path}/*.wav'))
@ -64,9 +65,10 @@ class MusicDiffusionFid(evaluator.Evaluator):
real_resampled = audio real_resampled = audio
audio = audio.unsqueeze(0) audio = audio.unsqueeze(0)
output_shape = audio.shape output_shape = audio.shape
gen = self.diffuser.p_sample_loop(self.model, output_shape, model_kwargs={'conditioning': audio, mel = self.spec_fn({'in': audio})['out']
'return_surrogate': False}) gen = self.diffuser.p_sample_loop(self.model, output_shape, noise=torch.zeros(*output_shape, device=audio.device),
#_, surrogate = self.model(audio, torch.tensor([0], device=audio.device), audio) model_kwargs={'aligned_conditioning': mel})
real_resampled = real_resampled + torch.FloatTensor(real_resampled.shape).uniform_(0.0, 1e-5).to(real_resampled.device)
return gen, real_resampled, sample_rate return gen, real_resampled, sample_rate
def load_projector(self): def load_projector(self):
@ -113,6 +115,10 @@ class MusicDiffusionFid(evaluator.Evaluator):
for i in tqdm(list(range(0, len(self.data), self.skip))): for i in tqdm(list(range(0, len(self.data), self.skip))):
path = self.data[i + self.env['rank']] path = self.data[i + self.env['rank']]
audio = load_audio(path, 22050).to(self.dev) audio = load_audio(path, 22050).to(self.dev)
mel = self.spec_fn({'in': audio})['out']
mel_norm = (mel + mel.min().abs())
mel_norm = mel_norm / mel_norm.max(dim=-1, keepdim=True).values
torchvision.utils.save_image(mel_norm.unsqueeze(1), 'mel.png')
sample, ref, sample_rate = self.diffusion_fn(audio) sample, ref, sample_rate = self.diffusion_fn(audio)
#gen_projections.append(self.project(projector, sample, sample_rate).cpu()) # Store on CPU to avoid wasting GPU memory. #gen_projections.append(self.project(projector, sample, sample_rate).cpu()) # Store on CPU to avoid wasting GPU memory.
@ -141,10 +147,10 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__': if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_waveform_gen2.yml', 'generator', diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_waveform_gen3.yml', 'generator',
also_load_savepoint=False, also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_waveform_gen2\\models\\59000_generator_ema.pth').cuda() load_path='X:\\dlas\\experiments\\train_music_waveform_gen3_r0\\models\\15400_generator_ema.pth').cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 500, opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100,
'conditioning_free': False, 'conditioning_free_k': 1, 'conditioning_free': False, 'conditioning_free_k': 1,
'diffusion_schedule': 'linear', 'diffusion_type': 'standard'} 'diffusion_schedule': 'linear', 'diffusion_type': 'standard'}
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 1, 'device': 'cuda', 'opt': {}} env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 1, 'device': 'cuda', 'opt': {}}