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

Restore causal decoding

Browse Source
This commit is contained in:
James Betker 2022-07-09 08:11:08 -06:00
parent 5138d61767
commit 99adc846c7
1 changed files with 6 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
codes/trainer/eval/music_diffusion_fid.py
View File

@ -253,7 +253,8 @@ class MusicDiffusionFid(evaluator.Evaluator):
cheater_codes = self.kmeans_inj({'in': cheater})['out'] cheater_codes = self.kmeans_inj({'in': cheater})['out']
ar_latent = self.local_modules['ar_prior'].to(audio.device)(cheater_codes, cheater, return_latent=True) ar_latent = self.local_modules['ar_prior'].to(audio.device)(cheater_codes, cheater, return_latent=True)
gen_mel = self.diffuser.ddim_sample_loop(self.model, mel_norm.shape, model_kwargs={'codes': ar_latent}, progress=True) gen_mel = self.diffuser.ddim_sample_loop(self.model, mel_norm.shape, model_kwargs={'codes': ar_latent},
causal=self.causal, causal_slope=self.causal_slope, progress=True)
gen_mel_denorm = denormalize_mel(gen_mel) gen_mel_denorm = denormalize_mel(gen_mel)
output_shape = (1,16,audio.shape[-1]//16) output_shape = (1,16,audio.shape[-1]//16)
@ -419,18 +420,18 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__': if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_cheater_gen_v5\\train.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\\models\\206000_generator_ema.pth' load_path='X:\\dlas\\experiments\\train_music_cheater_gen_v5_causal_retrain\\models\\12000_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': 64,
'conditioning_free': True, 'conditioning_free_k': 1, 'use_ddim': True, 'clip_audio': False, 'conditioning_free': True, 'conditioning_free_k': 1, 'use_ddim': True, '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': 4,
#'partial_low': 128, 'partial_high': 192 #'partial_low': 128, 'partial_high': 192
} }
env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 235, 'device': 'cuda', 'opt': {}} env = {'rank': 0, 'base_path': 'D:\\tmp\\test_eval_music', 'step': 236, 'device': 'cuda', 'opt': {}}
eval = MusicDiffusionFid(diffusion, opt_eval, env) eval = MusicDiffusionFid(diffusion, opt_eval, env)
print(eval.perform_eval()) print(eval.perform_eval())