mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

undo quantile sampler in music_diffusion_fid

Browse Source
This commit is contained in:
James Betker 2022-06-04 10:15:31 -06:00
parent a9387179db
commit 4819f15521
1 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
codes/trainer/eval/music_diffusion_fid.py
View File

@ -114,12 +114,12 @@ class MusicDiffusionFid(evaluator.Evaluator):
mel = self.spec_fn({'in': audio})['out']
mel_norm = normalize_mel(mel)
def denoising_fn(x):
q9 = torch.quantile(x, q=.95, dim=-1).unsqueeze(-1)
s = q9.clamp(1, 9999999999)
x = x.clamp(-s, s) / s
return x
gen_mel = self.diffuser.p_sample_loop(self.model, mel_norm.shape, denoised_fn=denoising_fn, clip_denoised=False,
#def denoising_fn(x):
# q9 = torch.quantile(x, q=.95, dim=-1).unsqueeze(-1)
# s = q9.clamp(1, 9999999999)
# x = x.clamp(-s, s) / s
# return x
gen_mel = self.diffuser.p_sample_loop(self.model, mel_norm.shape, #denoised_fn=denoising_fn, clip_denoised=False,
model_kwargs={'truth_mel': mel,
'conditioning_input': torch.zeros_like(mel_norm[:,:,:390]),
'disable_diversity': True})
@ -201,9 +201,9 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_diffusion_tfd5_quant\\train.yml', 'generator',
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_diffusion_tfd5_quant\\train_music_diffusion_tfd5_quant.yml', 'generator',
also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_diffusion_tfd5_quant\\models\\27000_generator_ema.pth'
load_path='X:\\dlas\\experiments\\train_music_diffusion_tfd5_quant\\models\\40500_generator_ema.pth'
).cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100,
'conditioning_free': True, 'conditioning_free_k': 1,