def pase_url(ckpt, model_config, refresh=False, **kwargs):
"""
The model from URL
ckpt (str): URL
model_config (str): URL
"""
ckpt = _urls_to_filepaths(ckpt, refresh=refresh)
model_config = _urls_to_filepaths(model_config, refresh=refresh)
return pase_local(ckpt, model_config, **kwargs)
def byol_a_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return byol_a_local(_urls_to_filepaths(ckpt, refresh=refresh), *args,
**kwargs)
def mockingjay_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return mockingjay_local(_urls_to_filepaths(ckpt, refresh=refresh), *args,
**kwargs)
def audio_albert_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return audio_albert_local(_urls_to_filepaths(ckpt, refresh=refresh), *args,
**kwargs)
def timit_posteriorgram_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return timit_posteriorgram_local(_urls_to_filepaths(ckpt, refresh=refresh),
*args, **kwargs)
def decoar_url(ckpt, refresh=False, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
ckpt = _urls_to_filepaths(ckpt, refresh=refresh)
return decoar_local(ckpt, **kwargs)
def wav2vec2_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from google drive id
ckpt (str): URL
refresh (bool): whether to download ckpt/config again if existed
"""
return wav2vec2_local(_urls_to_filepaths(ckpt, refresh=refresh), *args, **kwargs)
def spec_augment_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return spec_augment_local(_urls_to_filepaths(ckpt, refresh=refresh), *args,
**kwargs)
def vq_wav2vec_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from google drive id
ckpt (str): URL
feature_selection (str): z, codewords, codeids
refresh (bool): whether to download ckpt/config again if existed
"""
return vq_wav2vec_local(_urls_to_filepaths(ckpt, refresh=refresh), *args,
**kwargs)
def extract(files, args):
expdir = os.path.join(args.extract_dir, args.upstream)
os.makedirs(expdir, exist_ok=True)
model = getattr(hubconf, args.upstream)().to(device=args.device)
model.eval()
with torch.no_grad():
for file in files:
file_path = _urls_to_filepaths(file, refresh=args.refresh)
wav, sr = torchaudio.load(file_path)
wav = wav.view(-1).to(device=args.device)
repre = model([wav])[0].detach().cpu()
outpath = os.path.join(expdir, file.split('/')[-1] + '.pth')
torch.save(repre, outpath)
def test(files, args):
pths = [f'http://140.112.21.12:9000/extracted/{args.upstream}/' + file.split('/')[-1] + '.pth' for file in files]
model = getattr(hubconf, args.upstream)().to(device=args.device)
model.eval()
with torch.no_grad():
for file, pth in zip(files, pths):
file_path, pth_path = _urls_to_filepaths(file, pth, refresh=args.refresh)
wav, sr = torchaudio.load(file_path)
wav = wav.view(-1).to(device=args.device)
repre = model([wav])[0].detach().cpu()
max_diff = max((repre - torch.load(pth_path)).abs().max().item(), 0.0001)
with open(args.report, 'a') as handle:
handle.write(f'{max_diff} atol is required for torch.allclose to pass {args.upstream} + {file}.\n')
def vq_wav2vec_kmeans_roberta(refresh=False, **kwargs):
vq_wav2vec = getattr(hubconf, f'vq_wav2vec_kmeans')(refresh=refresh)
tar_file = _urls_to_filepaths(
'https://dl.fbaipublicfiles.com/fairseq/wav2vec/bert_kmeans.tar',
refresh=refresh)
tar_dir = os.path.join(os.path.dirname(tar_file),
'vq_wav2vec_kmeans_roberta/')
os.makedirs(tar_dir, exist_ok=True)
os.system(f'tar -xf {tar_file} -C {tar_dir}')
pt_files = glob.glob(os.path.join(tar_dir, '*.pt'))
assert len(pt_files) == 1
pt_file = pt_files[0]
kwargs['model_name_or_path'] = tar_dir
kwargs['checkpoint_file'] = pt_file
return _vq_wav2vec_roberta(vq_wav2vec, **kwargs)
def cpc_url(ckpt, refresh=False, *args, **kwargs):
"""
The model from URL
ckpt (str): URL
"""
return cpc_local(_urls_to_filepaths(ckpt), *args, **kwargs)
