kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

Move gen_similarities and rename

Browse Source
This commit is contained in:
James Betker 2022-03-16 11:59:34 -06:00
parent 8b376e63d9
commit 735f6e4640
1 changed files with 52 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
codes/scripts/audio/speech_to_speech_clip.py → codes/scripts/audio/preparation/phase3_generate_similarities.py
View File

@ -38,8 +38,10 @@ def process_subdir(subdir, options, clip_sz):
global clip_model global clip_model
if clip_model is None: if clip_model is None:
print('Loading CLIP model..') print('Loading CLIP model..')
clip_model = load_model_from_config(preloaded_options=options, model_name='clip', also_load_savepoint=True) clip_model = load_model_from_config(preloaded_options=options, model_name='clip', also_load_savepoint=True).cuda()
clip_model.eval()
with torch.no_grad():
root, paths = subdir root, paths = subdir
if len(paths) == 0: if len(paths) == 0:
return return
@ -63,7 +65,7 @@ def process_subdir(subdir, options, clip_sz):
while len(clips) > 0: while len(clips) > 0:
stacked = torch.stack(clips[:256], dim=0).cuda() stacked = torch.stack(clips[:256], dim=0).cuda()
clips = clips[256:] clips = clips[256:]
mels = wav_to_mel(stacked) mels = wav_to_mel(stacked).cuda()
outp = clip_model.inference(mels).cpu() outp = clip_model.inference(mels).cpu()
if sims is None: if sims is None:
sims = outp sims = outp
@ -94,12 +96,13 @@ if __name__ == '__main__':
""" """
This script iterates within a directory filled with subdirs. Each subdir contains a list of audio files from the same This script iterates within a directory filled with subdirs. Each subdir contains a list of audio files from the same
source. The script uses an speech-to-speech clip model to find the <n> most similar audio clips within each subdir for source. The script uses an speech-to-speech clip model to find the <n> most similar audio clips within each subdir for
each clip within that subdir. each clip within that subdir. These similar files are recorded in a "similarities.pth" file in each subdirectory, which
is consumed during training when the dataset searches for conditioning clips.
""" """
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('-o', type=str, help='Path to the options YAML file used to train the CLIP model', default='../options/train_voice_voice_clip.yml') parser.add_argument('-o', type=str, help='Path to the options YAML file used to train the CLIP model', default='../options/train_voice_voice_clip.yml')
parser.add_argument('--num_workers', type=int, help='Number concurrent processes to use', default=6) parser.add_argument('--num_workers', type=int, help='Number concurrent processes to use', default=6)
parser.add_argument('--root_path', type=str, help='Root path to search for audio directories from', default='Y:\\bigasr_dataset\\tedlium') parser.add_argument('--root_path', type=str, help='Root path to search for audio directories from', default='Y:\\filtered\\big_podcast')
parser.add_argument('--clip_size', type=int, help='Amount of audio samples to pull from each file', default=22050) parser.add_argument('--clip_size', type=int, help='Amount of audio samples to pull from each file', default=22050)
args = parser.parse_args() args = parser.parse_args()