config = M1.get_default_config()
net = M1.SpeechTranscriber(config)
run_config = dict(max_norm=2.0, max_lr=2 * 1e-4, epochs=32, opt='adam')
logging.info('Training ASR')
M1.experiment(net, data, run_config)
suffix = str(ds_factor).zfill(lz)
res_fname = 'result_asr_{}.json'.format(suffix)
copyfile('result.json', res_fname)
net_fname = 'asr_{}.best.pt'.format(ds_factor)
copy_best(res_fname, net_fname, experiment_type='asr')
net = torch.load(net_fname)
logging.info('Extracting ASR transcriptions')
for set_name in ['train', 'val']:
ds = data[set_name].dataset
hyp_asr, ref_asr = extract_trn(net, ds, use_beam_decoding=True)
# Replacing original transcriptions with ASR's output
for i in range(len(hyp_asr)):
item = ds.split_data[i]
if item[2] == ref_asr[i]:
ds.split_data[i] = (item[0], item[1], hyp_asr[i])
else:
msg = 'Extracted reference #{} ({}) doesn\'t match dataset\'s \
one ({}) for {} set.'
msg = msg.format(i, ref_asr[i], ds.split_data[i][3], set_name)
logging.warning(msg)
if args.asr_model_dir:
# Saving config for text-image model
pickle.dump(
else:
logging.info('Building ASR model')
config = M1.get_default_config()
net = M1.SpeechTranscriber(config)
run_config = dict(max_norm=2.0, max_lr=2 * 1e-4, epochs=32, opt='adam')
logging.info('Training ASR')
M1.experiment(net, data, run_config)
suffix = str(ds_factor).zfill(lz)
res_fname = 'result_asr_{}.json'.format(suffix)
copyfile('result.json', res_fname)
net_fname = 'asr_{}.best.pt'.format(ds_factor)
copy_best(res_fname, net_fname, experiment_type='asr')
net = torch.load(net_fname)
logging.info('Extracting ASR transcriptions')
hyp_asr, _ = extract_trn(net, data['val'].dataset, use_beam_decoding=True)
if args.text_image_model_dir:
config_fpath = os.path.join(args.text_image_model_dir, 'config.pkl')
config = pickle.load(open(config_fpath, 'rb'))
fd.le = config['label_encoder']
elif args.asr_model_dir:
# Saving config for text-image model
pickle.dump(dict(feature_fname=feature_fname,
label_encoder=fd.get_label_encoder(),
language='en'),
open('config.pkl', 'wb'))
if args.text_image_model_dir:
net_fname = 'net_{}.best.pt'.format(ds_factor)
net = torch.load(os.path.join(args.text_image_model_dir, net_fname))