def h5features_stack_fbanks(fbanks_file, stacked_fbanks_file):
import h5features
index = h5features.read_index(fbanks_file)
files = index['files']
for f in files:
times, fbanks = h5features.read(
fbanks_file, 'features', from_internal_file=f, index=index)
stacked_fbanks = stack_fbanks(fbanks[f])
h5features.write(stacked_fbanks_file, 'features', [f],
[times[f]], [stacked_fbanks])
def decode_from_features(model_path,
wavefile_list,
output_file,
features_file,
verbose=0):
"""Function to calculate the embeddings for a list of wavefiles using
a trained ABnet in the h5features format
Parameters:
-----------
model_path: string, path to the abnet
wavefile_list: string, path to the list of wavefiles (wavefiles
are specified by their full path, one file per line)
output_file: string, path to the outputfile, it will contains the
embeddings in h5features format
"""
with open(wavefile_list) as fin:
wavefiles = [w.strip() for w in fin]
nnet_file = p.join(model_path, 'model.pickle')
mean_std_file = p.join(model_path, 'pairs_mean_std.npz')
stacked_fb_file = p.join(model_path, 'stackedfeats.h5f')
prepare.h5features_feats2stackedfeats(features_file, stacked_fb_file)
if verbose:
print('stacked fbanks calculated')
with open(nnet_file, 'rb') as f:
nnet = cPickle.load(f)
with open(p.join(model_path, 'nframes')) as fin:
NFRAMES = int(fin.readline().strip())
transform = nnet.transform_x1()
tmp = np.load(mean_std_file)
mean = np.tile(tmp['mean'], NFRAMES)
std = np.tile(tmp['std'], NFRAMES)
if verbose:
print('model loaded')
index = h5features.read_index(stacked_fb_file)
# TODO maybe normalize embedded features ???
def do_transform(fname):
times, feats = h5features.read(stacked_fb_file,
from_internal_file=fname,
index=index)
times = times[fname]
feats = feats[fname]
X = np.asarray((feats - mean) / std, dtype='float32')
# times = np.arange(0.01, 0.01*npz.shape[0], 0.01)
emb_wrd, emb_spkr = transform(X)
return emb_wrd
prepare.h5features_fbanks(index['files'],
output_file,
featfunc=do_transform)
if verbose:
print('bottleneck features calculated, all done.')
def __init__(self, feature_file):
self.feature_file = feature_file
self.index = h5features.read_index(feature_file)
