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

Update use_diffuse_tts

Browse Source
This commit is contained in:
James Betker 2022-01-27 19:57:28 -07:00
parent a77d376ad2
commit e0e36ed98c
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/scripts/audio/gen/use_diffuse_tts.py
View File

@ -22,13 +22,15 @@ if __name__ == '__main__':
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts_medium\\train_diffusion_tts_medium.yml') parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='X:\\dlas\\experiments\\train_diffusion_tts_medium\\train_diffusion_tts_medium.yml')
parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator') parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator')
parser.add_argument('-diffusion_model_path', type=str, help='Path to saved model weights', default='X:\\dlas\\experiments\\train_diffusion_tts_medium\\models\\22500_generator_ema.pth') parser.add_argument('-diffusion_model_path', type=str, help='Path to saved model weights', default='X:\\dlas\\experiments\\train_diffusion_tts_medium\\models\\38500_generator_ema.pth')
parser.add_argument('-aligned_codes', type=str, help='Comma-delimited list of integer codes that defines text & prosody. Get this by apply W2V to an existing audio clip or from a bespoke generator.', parser.add_argument('-aligned_codes', type=str, help='Comma-delimited list of integer codes that defines text & prosody. Get this by apply W2V to an existing audio clip or from a bespoke generator.',
default='0,0,0,0,10,10,0,4,0,7,0,17,4,4,0,25,5,0,13,13,0,22,4,4,0,21,15,15,7,0,0,14,4,4,6,8,4,4,0,0,12,5,0,0,5,0,4,4,22,22,8,16,16,0,4,4,4,0,0,0,0,0,0,0') # Default: 'i am very glad to see you', libritts/train-clean-100/103/1241/103_1241_000017_000001.wav. default='0,0,0,0,10,10,0,4,0,7,0,17,4,4,0,25,5,0,13,13,0,22,4,4,0,21,15,15,7,0,0,14,4,4,6,8,4,4,0,0,12,5,0,0,5,0,4,4,22,22,8,16,16,0,4,4,4,0,0,0,0,0,0,0') # Default: 'i am very glad to see you', libritts/train-clean-100/103/1241/103_1241_000017_000001.wav.
# -cond "Y:\libritts/train-clean-100/103/1241/103_1241_000017_000001.wav" # -cond "Y:\libritts/train-clean-100/103/1241/103_1241_000017_000001.wav"
parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav') parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav')
parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100) parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100)
parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_tts') parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_tts')
parser.add_argument('-sample_rate', type=int, help='Model sample rate', default=11025)
parser.add_argument('-cond_sample_rate', type=int, help='Model sample rate', default=22050)
parser.add_argument('-device', type=str, help='Device to run on', default='cpu') parser.add_argument('-device', type=str, help='Device to run on', default='cpu')
args = parser.parse_args() args = parser.parse_args()
os.makedirs(args.output_path, exist_ok=True) os.makedirs(args.output_path, exist_ok=True)
@ -36,12 +38,12 @@ if __name__ == '__main__':
print("Loading Diffusion Model..") print("Loading Diffusion Model..")
diffusion = load_model_from_config(args.opt, args.diffusion_model_name, also_load_savepoint=False, diffusion = load_model_from_config(args.opt, args.diffusion_model_name, also_load_savepoint=False,
load_path=args.diffusion_model_path, device=args.device) load_path=args.diffusion_model_path, device=args.device)
aligned_codes_compression_factor = 221 # Derived empirically for 11025Hz sample rate. Adjust if sample rate increases. aligned_codes_compression_factor = args.sample_rate * 221 // 11025
print("Loading data..") print("Loading data..")
aligned_codes = torch.tensor([int(s) for s in args.aligned_codes.split(',')]).to(args.device) aligned_codes = torch.tensor([int(s) for s in args.aligned_codes.split(',')]).to(args.device)
diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=args.diffusion_steps) diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=args.diffusion_steps)
cond = load_audio(args.cond, 22050).to(args.device) cond = load_audio(args.cond, args.cond_sample_rate).to(args.device)
if cond.shape[-1] > 88000: if cond.shape[-1] > 88000:
cond = cond[:,:88000] cond = cond[:,:88000]
@ -53,10 +55,10 @@ if __name__ == '__main__':
output = diffuser.p_sample_loop(diffusion, output_shape, noise=torch.zeros(output_shape, device=args.device), output = diffuser.p_sample_loop(diffusion, output_shape, noise=torch.zeros(output_shape, device=args.device),
model_kwargs={'tokens': aligned_codes.unsqueeze(0), model_kwargs={'tokens': aligned_codes.unsqueeze(0),
'conditioning_input': cond.unsqueeze(0)}) 'conditioning_input': cond.unsqueeze(0)})
torchaudio.save(os.path.join(args.output_path, f'output_mean.wav'), output.cpu().squeeze(0), 11025) torchaudio.save(os.path.join(args.output_path, f'output_mean.wav'), output.cpu().squeeze(0), args.sample_rate)
for k in range(5): for k in range(5):
output = diffuser.p_sample_loop(diffusion, output_shape, model_kwargs={'tokens': aligned_codes.unsqueeze(0), output = diffuser.p_sample_loop(diffusion, output_shape, model_kwargs={'tokens': aligned_codes.unsqueeze(0),
'conditioning_input': cond.unsqueeze(0)}) 'conditioning_input': cond.unsqueeze(0)})
torchaudio.save(os.path.join(args.output_path, f'output_{k}.wav'), output.cpu().squeeze(0), 11025) torchaudio.save(os.path.join(args.output_path, f'output_{k}.wav'), output.cpu().squeeze(0), args.sample_rate)