num_hidden_proj=num_hidden_proj)
data_names = [x[0] for x in data_test.provide_data]
label_names = ['softmax_label']
module = mx.mod.Module(sym,
context=contexts,
data_names=data_names,
label_names=label_names)
# set the parameters
module.bind(data_shapes=data_test.provide_data,
label_shapes=None,
for_training=False)
module.set_params(arg_params=arg_params, aux_params=aux_params)
#kaldiWriter = KaldiWriteOut(None, out_file)
#kaldiWriter.open_or_fd()
kaldiWriter = KaldiWriteOut(out_dir + "/feats.scp", out_dir + "/feats.ark")
kaldiWriter.open()
for preds, i_batch, batch in module.iter_predict(data_test):
label = batch.label[0].asnumpy().astype('int32')
posteriors = preds[0].asnumpy().astype('float32')
# copy over states
if decoding_method == METHOD_BUCKETING:
for (ind, utt) in enumerate(batch.utt_id):
if utt != "GAP_UTT":
posteriors = np.log(posteriors[:label[0][0], 1:] +
1e-20) - np.log(
data_train.label_mean).T
kaldiWriter.write(utt, posteriors)
elif decoding_method == METHOD_SIMPLE:
for (ind, utt) in enumerate(batch.utt_id):
if utt != "GAP_UTT":
num_label=label_dim,
output_states=True,
num_hidden_proj=num_hidden_proj)
data_names = [x[0] for x in data_test.provide_data]
label_names = ['softmax_label']
module = mx.mod.Module(sym,
context=contexts,
data_names=data_names,
label_names=label_names)
# set the parameters
module.bind(data_shapes=data_test.provide_data,
label_shapes=None,
for_training=False)
module.set_params(arg_params=arg_params, aux_params=aux_params)
kaldiWriter = KaldiWriteOut(None, out_file)
kaldiWriter.open_or_fd()
for preds, i_batch, batch in module.iter_predict(data_test):
#pred_label = np.array(preds[0].asnumpy().argmax(axis=1))
label = batch.label[0].asnumpy().astype('int32')
posteriors = preds[0].asnumpy().astype('float32')[0]
#print np.sum(posteriors[1][:])
# copy over states
if decoding_method == METHOD_BUCKETING:
for (ind, utt) in enumerate(batch.utt_id):
if utt != "GAP_UTT":
#print sum(posteriors[0,:])
posteriors = np.log(posteriors[:label[0][0], 1:] +
1e-20) - np.log(
data_train.label_mean).T
truncate_len=20
data_test = TruncatedSentenceIter(test_sets, batch_size, init_states,
truncate_len, feat_dim=feat_dim,
do_shuffling=False, pad_zeros=True, has_label=True)
sym = lstm_unroll(num_lstm_layer, truncate_len, feat_dim, num_hidden=num_hidden,
num_label=label_dim, output_states=True, num_hidden_proj=num_hidden_proj)
data_names = [x[0] for x in data_test.provide_data]
label_names = ['softmax_label']
module = mx.mod.Module(sym, context=contexts, data_names=data_names,
label_names=label_names)
# set the parameters
module.bind(data_shapes=data_test.provide_data, label_shapes=None, for_training=False)
module.set_params(arg_params=arg_params, aux_params=aux_params)
kaldiWriter = KaldiWriteOut(None, out_file)
kaldiWriter.open_or_fd()
for preds, i_batch, batch in module.iter_predict(data_test):
#pred_label = np.array(preds[0].asnumpy().argmax(axis=1))
label = batch.label[0].asnumpy().astype('int32')
posteriors = preds[0].asnumpy().astype('float32')[0]
#print np.sum(posteriors[1][:])
# copy over states
if decoding_method == METHOD_BUCKETING:
for (ind, utt) in enumerate(batch.utt_id):
if utt != "GAP_UTT":
#print sum(posteriors[0,:])
posteriors = np.log(posteriors[:label[0][0],1:] + 1e-20) - np.log(data_train.label_mean).T
kaldiWriter.write(utt, posteriors)
elif decoding_method == METHOD_SIMPLE:
init_states, test_sets = prepare_data(args)
state_names = [x[0] for x in init_states]
batch_size = args.config.getint('train', 'batch_size')
num_hidden = args.config.getint('arch', 'num_hidden')
num_lstm_layer = args.config.getint('arch', 'num_lstm_layer')
feat_dim = args.config.getint('data', 'xdim')
label_dim = args.config.getint('data', 'ydim')
out_file = args.config.get('data', 'out_file')
num_epoch = args.config.getint('train', 'num_epoch')
model_name = get_checkpoint_path(args)
logging.basicConfig(level=logging.DEBUG, format='%(asctime)-15s %(message)s')
# load the model
label_mean = np.zeros((label_dim,1), dtype='float32')
data_test = TruncatedSentenceIter(test_sets, batch_size, init_states,
20, feat_dim=feat_dim,
do_shuffling=False, pad_zeros=True, has_label=True)
for i, batch in enumerate(data_test.labels):
hist, edges = np.histogram(batch.flat, bins=range(0,label_dim+1))
label_mean += hist.reshape(label_dim,1)
kaldiWriter = KaldiWriteOut(None, out_file)
kaldiWriter.open_or_fd()
kaldiWriter.write("label_mean", label_mean)
args.config.write(sys.stderr)
num_hidden = args.config.getint('arch', 'num_hidden')
num_lstm_layer = args.config.getint('arch', 'num_lstm_layer')
feat_dim = args.config.getint('data', 'xdim')
label_dim = args.config.getint('data', 'ydim')
out_file = args.config.get('data', 'out_file')
num_epoch = args.config.getint('train', 'num_epoch')
model_name = get_checkpoint_path(args)
logging.basicConfig(level=logging.DEBUG,
format='%(asctime)-15s %(message)s')
# load the model
label_mean = np.zeros((label_dim, 1), dtype='float32')
data_test = TruncatedSentenceIter(test_sets,
batch_size,
init_states,
20,
feat_dim=feat_dim,
do_shuffling=False,
pad_zeros=True,
has_label=True)
for i, batch in enumerate(data_test.labels):
hist, edges = np.histogram(batch.flat, bins=range(0, label_dim + 1))
label_mean += hist.reshape(label_dim, 1)
kaldiWriter = KaldiWriteOut(None, out_file)
kaldiWriter.open_or_fd()
kaldiWriter.write("label_mean", label_mean)
args.config.write(sys.stderr)
