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

Constrain wav_aug a bit more

Browse Source
This commit is contained in:
James Betker 2021-08-06 08:19:38 -06:00
parent f126040da2
commit 62c7570512
3 changed files with 9 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
codes/data/audio/wav_aug.py
View File

@ -23,7 +23,9 @@ class WavAugmentor:
pass pass
def augment(self, wav, sample_rate): def augment(self, wav, sample_rate):
speed_effect = ['speed', rdstr(.7, 1)] speed_effect = ['speed', rdstr(.8, 1)]
'''
Band effects are disabled until I can audit them better.
band_effects = [ band_effects = [
['reverb', '-w'], ['reverb', '-w'],
['reverb'], ['reverb'],
@ -39,15 +41,16 @@ class WavAugmentor:
['sinc', '3k-4k'] ['sinc', '3k-4k']
] ]
band_effect = random.choice(band_effects) band_effect = random.choice(band_effects)
'''
volume_effects = [ volume_effects = [
['loudness', rdi(10,-2)], ['loudness', rdi(10,-2)],
['overdrive', rdi(20,0), rdi(20,0)], ['overdrive', rdi(20,0), rdi(20,0)],
] ]
vol_effect = random.choice(volume_effects) vol_effect = random.choice(volume_effects)
effects = [speed_effect, band_effect, vol_effect] effects = [speed_effect, vol_effect]
out, sr = torchaudio.sox_effects.apply_effects_tensor(wav, sample_rate, effects) out, sr = torchaudio.sox_effects.apply_effects_tensor(wav, sample_rate, effects)
# Add a variable amount of noise # Add a variable amount of noise
out = out + torch.rand_like(out) * random.random() * .05 out = out + torch.rand_like(out) * random.random() * .03
return out return out

2
codes/scripts/audio/test_audio_gen.py
View File

@ -51,7 +51,7 @@ if __name__ == "__main__":
torch.backends.cudnn.benchmark = True torch.backends.cudnn.benchmark = True
want_metrics = False want_metrics = False
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_vqvae_audio_lj.yml') parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_gpt_tts_lj.yml')
opt = option.parse(parser.parse_args().opt, is_train=False) opt = option.parse(parser.parse_args().opt, is_train=False)
opt = option.dict_to_nonedict(opt) opt = option.dict_to_nonedict(opt)
utils.util.loaded_options = opt utils.util.loaded_options = opt

3
codes/scripts/audio/test_audio_similarity.py
View File

@ -20,12 +20,13 @@ if __name__ == '__main__':
clip = clip[:,0] clip = clip[:,0]
clip = clip[:window].unsqueeze(0) clip = clip[:window].unsqueeze(0)
clip = clip / 32768.0 # Normalize clip = clip / 32768.0 # Normalize
clip = clip + torch.rand_like(clip) * .03 # Noise (this is how the model was trained)
assert sr == 24000 assert sr == 24000
clips.append(clip) clips.append(clip)
clips = torch.stack(clips, dim=0) clips = torch.stack(clips, dim=0)
resnet = resnet34() resnet = resnet34()
sd = torch.load('../experiments/train_byol_audio_clips/models/66000_generator.pth') sd = torch.load('../experiments/train_byol_audio_clips/models/57000_generator.pth')
sd = extract_byol_model_from_state_dict(sd) sd = extract_byol_model_from_state_dict(sd)
resnet.load_state_dict(sd) resnet.load_state_dict(sd)
embedding = resnet(clips, return_pool=True) embedding = resnet(clips, return_pool=True)