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

precompute diffusion embeddings for from_codes

Browse Source
This commit is contained in:
James Betker 2022-05-22 06:45:57 -06:00
parent ea21a8b107
commit db38672dae
2 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
codes/trainer/eval/music_diffusion_fid.py
View File

@ -167,7 +167,10 @@ class MusicDiffusionFid(evaluator.Evaluator):
codegen = self.local_modules['codegen'].to(mel.device)
codes = codegen.get_codes(mel)
mel_norm = normalize_mel(mel)
gen_mel = self.diffuser.p_sample_loop(self.model, mel_norm.shape, model_kwargs={'aligned_conditioning': codes, 'conditioning_input': mel[:,:,:117]})
precomputed = self.model.timestep_independent(aligned_conditioning=codes, conditioning_input=mel[:,:,:112],
expected_seq_len=mel_norm.shape[-1], return_code_pred=False)
gen_mel = self.diffuser.p_sample_loop(self.model, mel_norm.shape, noise=torch.zeros_like(mel_norm),
model_kwargs={'precomputed_aligned_embeddings': precomputed})
gen_mel_denorm = denormalize_mel(gen_mel)
output_shape = (1,16,audio.shape[-1]//16)
@ -243,7 +246,7 @@ class MusicDiffusionFid(evaluator.Evaluator):
if __name__ == '__main__':
diffusion = load_model_from_config('X:\\dlas\\experiments\\train_music_diffusion_flat.yml', 'generator',
also_load_savepoint=False,
load_path='X:\\dlas\\experiments\\train_music_diffusion_flat\\models\\26000_generator.pth').cuda()
load_path='X:\\dlas\\experiments\\train_music_diffusion_flat\\models\\33000_generator_ema.pth').cuda()
opt_eval = {'path': 'Y:\\split\\yt-music-eval', 'diffusion_steps': 100,
'conditioning_free': False, 'conditioning_free_k': 1,
'diffusion_schedule': 'linear', 'diffusion_type': 'from_codes'}

1
codes/utils/util.py
View File

@ -561,6 +561,7 @@ def find_audio_files(base_path, globs=['*.wav', '*.mp3', '*.ogg', '*.flac']):
def load_audio(audiopath, sampling_rate, raw_data=None):
audiopath = str(audiopath)
if raw_data is not None:
# Assume the data is wav format. SciPy's reader can read raw WAV data from a BytesIO wrapper.
audio, lsr = load_wav_to_torch(raw_data)