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

music diffusion fid adjustments

Browse Source
This commit is contained in:
James Betker 2022-04-28 10:08:55 -06:00
parent ab8176b217
commit 3f67cb2023
1 changed files with 7 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
codes/trainer/eval/music_diffusion_fid.py
View File

@ -62,9 +62,11 @@ class MusicDiffusionFid(evaluator.Evaluator):
real_resampled = torchaudio.functional.resample(audio, 22050, sample_rate).unsqueeze(0)
else:
real_resampled = audio
mel = self.spec_fn({'in': real_resampled})['out']
output_shape = (1, 1, mel.shape[-1] * 256)
gen = self.diffuser.p_sample_loop(self.model, output_shape, model_kwargs={'aligned_conditioning': mel})
audio = audio.unsqueeze(0)
output_shape = audio.shape
gen = self.diffuser.p_sample_loop(self.model, output_shape, model_kwargs={'conditioning': audio,
'return_surrogate': False})
#_, surrogate = self.model(audio, torch.tensor([0], device=audio.device), audio)
return gen, real_resampled, sample_rate
def load_projector(self):
@ -139,9 +141,9 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_waveform_gen.yml', 'generator',
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_waveform_gen2.yml', 'generator',
also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_waveform_gen_r3\\models\\11200_generator_ema.pth').cuda()
load_path='X:\\dlas\\experiments\\train_music_waveform_gen2\\models\\59000_generator_ema.pth').cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 500,
'conditioning_free': False, 'conditioning_free_k': 1,
'diffusion_schedule': 'linear', 'diffusion_type': 'standard'}