kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

mdf: re-allow p_sample_loop

Browse Source
This commit is contained in:
James Betker 2022-07-09 15:35:09 -06:00
parent f28cefdfe2
commit b432d7c7de
1 changed files with 9 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
codes/trainer/eval/music_diffusion_fid.py
View File

@ -215,7 +215,7 @@ class MusicDiffusionFid(evaluator.Evaluator):
return gen_wav, real_resampled, gen_mel, mel_norm, sample_rate return gen_wav, real_resampled, gen_mel, mel_norm, sample_rate
def perform_reconstruction_from_cheater_gen(self, audio, sample_rate=22050): def perform_reconstruction_from_cheater_gen(self, audio, sample_rate=22050):
assert self.ddim, "DDIM mode expected for reconstructing cheater gen. Do you like to waste resources??" #assert self.ddim, "DDIM mode expected for reconstructing cheater gen. Do you like to waste resources??"
audio = audio.unsqueeze(0) audio = audio.unsqueeze(0)
mel = self.spec_fn({'in': audio})['out'] mel = self.spec_fn({'in': audio})['out']
@ -223,7 +223,8 @@ class MusicDiffusionFid(evaluator.Evaluator):
cheater = self.local_modules['cheater_encoder'].to(audio.device)(mel_norm) cheater = self.local_modules['cheater_encoder'].to(audio.device)(mel_norm)
# 1. Generate the cheater latent using the input as a reference. # 1. Generate the cheater latent using the input as a reference.
gen_cheater = self.diffuser.ddim_sample_loop(self.model, cheater.shape, progress=True, sampler = self.diffuser.ddim_sample_loop if self.ddim else self.diffuser.p_sample_loop
gen_cheater = sampler(self.model, cheater.shape, progress=True,
causal=self.causal, causal_slope=self.causal_slope, causal=self.causal, causal_slope=self.causal_slope,
model_kwargs={'conditioning_input': cheater}) model_kwargs={'conditioning_input': cheater})
@ -423,14 +424,14 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__': if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_cheater_gen.yml', 'generator', diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_cheater_gen.yml', 'generator',
also_load_savepoint=False, also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_cheater_gen_v5_causal_retrain\\models\\12000_generator.pth' load_path='X:\\dlas\\experiments\\train_music_cheater_gen_v5_causal_retrain\\models\\18000_generator.pth'
).cuda() ).cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', # eval music, mostly electronica. :) opt_eval = {'path': 'Y:\\split\\yt-music-eval', # eval music, mostly electronica. :)
#'path': 'E:\\music_eval', # this is music from the training dataset, including a lot more variety. #'path': 'E:\\music_eval', # this is music from the training dataset, including a lot more variety.
'diffusion_steps': 64, 'diffusion_steps': 256,
'conditioning_free': True, 'conditioning_free_k': 1, 'use_ddim': True, 'clip_audio': False, 'conditioning_free': True, 'conditioning_free_k': 1, 'use_ddim': False, 'clip_audio': False,
'diffusion_schedule': 'linear', 'diffusion_type': 'cheater_gen', 'diffusion_schedule': 'linear', 'diffusion_type': 'cheater_gen',
'causal': True, 'causal_slope': 4, 'causal': True, 'causal_slope': 1,
#'partial_low': 128, 'partial_high': 192 #'partial_low': 128, 'partial_high': 192
} }
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 236, 'device': 'cuda', 'opt': {}} env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 236, 'device': 'cuda', 'opt': {}}