import argparse
from dnnseg.config import Config
from dnnseg.util import load_dnnseg, stderr
if __name__ == '__main__':
argparser = argparse.ArgumentParser('''
Trains a DNN-Seg model from a config file.
''')
argparser.add_argument('config', help='Path to configuration file.')
argparser.add_argument('-p', '--partition', default='train', help='Name of partition from which to extract classifications (one of ["train", "dev", "test"]')
argparser.add_argument('-s', '--segtype', type=str, default='phn', help='Segment type to use for training (one of ["vad", "wrd", "phn"]')
argparser.add_argument('-g', '--goldfeats', type=str, default=None, help='Path to gold features file to merge with classification table. If unspecified, no gold features columns will be added.')
args = argparser.parse_args()
p = Config(args.config)
if not p['use_gpu_if_available']:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
t0 = time.time()
data_path = 'data'
if p['os']:
data_path += '_os'
if p['root']:
data_path += '_root'
data_path += '.obj'
stderr('Loading saved training data...\n')
sys.stderr.flush()
'--force_cpu',
action='store_true',
help=
'Do not use GPU. If not specified, GPU usage defaults to the value of the **use_gpu_if_available** configuration parameter.'
)
argparser.add_argument('-o',
'--outdir',
default=None,
help='Directory in which to save outputs.')
args = argparser.parse_args()
outputs = {}
val_data = None
for config in args.config:
p = Config(config)
if args.force_cpu or not p['use_gpu_if_available']:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
t0 = time.time()
if p['data_type'].lower() == 'acoustic':
data_name = 'data_%s_f%s_d%s.obj' % (p['filter_type'], p['n_coef'],
p['order'])
else:
data_name = 'data.obj'
preprocessed = True
if val_data is None:
'-l',
'--layers',
default=None,
nargs='+',
help=
'IDs of layers to plot (0, ..., L). If unspecified, plots all available layers.'
)
args = argparser.parse_args()
if args.class_limit.lower() == 'inf':
class_limit = np.inf
else:
class_limit = int(args.class_limit)
for config_path in args.config:
p = Config(config_path)
stderr('Plotting model defined at path %s...\n' % p['outdir'])
if p['label_map_file'] is not None and os.path.exists(
p['label_map_file']):
label_map = pd.read_csv(p['label_map_file'])
label_map = dict(zip(label_map.source, label_map.target))
else:
label_map = None
if p['feature_map_file'] is not None and os.path.exists(
p['feature_map_file']):
feature_table = pd.read_csv(p['feature_map_file'])
new_cols = []
for i in range(len(feature_table.columns)):
new_cols.append(feature_table.columns[i].replace(' ',
'Do not use GPU. If not specified, GPU usage defaults to the value of the **use_gpu_if_available** configuration parameter.'
)
args = argparser.parse_args()
measures = []
for m in args.eval_measures:
if m.lower() == 'all':
measures = [
"all", "segmentation", "classification", "objective",
"label_probe", "feature_probe"
]
break
measures.append(m.lower())
for c in args.configs:
p = Config(c)
if args.force_cpu or not p['use_gpu_if_available']:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
t0 = time.time()
if 'segmentation' in measures or 'classification' in measures or 'objective' in measures or args.force_predict:
if p['data_type'].lower() == 'acoustic':
data_name = 'data_%s_f%s_d%s.obj' % (p['filter_type'],
p['n_coef'], p['order'])
else:
data_name = 'data.obj'
if 'objective' in measures:
train_data_dirs = p.train_data_dir.split()