split='train', batch_size=batch_size, shuffle=True,
downsampling_factor=ds_factor_text),
val=D.flickr8k_loader(split='val', batch_size=batch_size))
else:
data_asr = data
if args.asr_model_dir:
net = torch.load(os.path.join(args.asr_model_dir, 'net.best.pt'))
else:
logging.info('Building ASR/SLT model')
config = M1.get_default_config()
net = M1.SpeechTranscriber(config)
run_config = dict(max_norm=2.0, max_lr=2 * 1e-4, epochs=32)
logging.info('Training ASR/SLT')
if data_asr['train'].dataset.is_slt():
M1.experiment(net, data_asr, run_config, slt=True)
copy_best('.', 'result.json', 'asr.best.pt', experiment_type='slt')
else:
M1.experiment(net, data_asr, run_config)
copy_best('.', 'result.json', 'asr.best.pt', experiment_type='asr')
copyfile('result.json', 'result_asr.json')
net = torch.load('asr.best.pt')
logging.info('Extracting ASR/SLT 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/SLT's output
for i in range(len(hyp_asr)):
item = ds.split_data[i]
if item[2] == ref_asr[i]:
num_layers=6,
bidirectional=True,
dropout=dropout),
rnn_layer_type=nn.GRU),
TextDecoder=dict(
emb=dict(num_embeddings=fd.vocabulary_size(),
embedding_dim=hidden_size),
drop=dict(p=dropout),
att=dict(in_size_enc=hidden_size * 2,
in_size_state=hidden_size,
hidden_size=hidden_size),
rnn=dict(input_size=hidden_size * 3,
hidden_size=hidden_size,
num_layers=1,
dropout=dropout),
out=dict(in_features=hidden_size * 3,
out_features=fd.vocabulary_size()),
rnn_layer_type=nn.GRU,
max_output_length=400, # max length for flickr annotations is 199
sos_id=fd.get_token_id(fd.sos),
eos_id=fd.get_token_id(fd.eos),
pad_id=fd.get_token_id(fd.pad)),
inverse_transform_fn=fd.get_label_encoder().inverse_transform)
logging.info('Building model')
net = M.SpeechTranscriber(config)
run_config = dict(max_norm=2.0, max_lr=2 * 1e-4, epochs=32, opt='adam')
logging.info('Training')
M.experiment(net, data, run_config)
args.enable_help()
args.parse()
# Setting general configuration
torch.manual_seed(args.seed)
random.seed(args.seed)
# Logging the arguments
logging.info('Arguments: {}'.format(args))
batch_size = 8
logging.info('Loading data')
data = dict(
train=D.flickr8k_loader(
args.flickr8k_root, args.flickr8k_meta, args.flickr8k_language,
args.audio_features_fn, split='train', batch_size=batch_size,
shuffle=True, downsampling_factor=args.downsampling_factor),
val=D.flickr8k_loader(
args.flickr8k_root, args.flickr8k_meta, args.flickr8k_language,
args.audio_features_fn, split='val', batch_size=batch_size,
shuffle=False))
logging.info('Building model')
net = M.SpeechTranscriber(M.get_default_config())
run_config = dict(max_norm=2.0, max_lr=2 * 1e-4, epochs=args.epochs)
logging.info('Training')
M.experiment(net, data, run_config, slt=data['train'].dataset.is_slt())
