def prediction(args,
number,
data_dir='../samples/cha/stimuli3/0',
model_path="",
mode='train',
data_suffix='npy'):
""" Get the ground truth and predicted sequence
number: number of data in the batch wants to be visualized
mode: train or validation or test, which will indicate the folder to be opened
dir: the directory of the mfcc data
model_path: the model checkpoint, HDF5 file
data_suffix: the data suffix of those audio file inside the data_dir
"""
# get data by index
data, label, total_path_name = data_specification(mode, data_dir,
data_suffix)
max_sequence_length = 0
for i in data:
max_sequence_length = max(max_sequence_length, i.shape[1])
# Training Phase
model = brnn_keras(args, max_sequence_length)
#model.summary()
model.model_1.load_weights(model_path)
#model.model_2.load_weights(model_path)
batch = create_batch(args, data, max_sequence_length, label)
for i in range(number):
inputs, outputs = next(batch)
x = inputs['the_input']
y = inputs['the_labels'][0]
prediction = model.model_1.predict(x, steps=1)
pred_ints = decode_ctc(args, prediction)
print(np.shape(prediction))
model_path = model_path.replace("../CHECKPOINT/char/save/", "")
print('model ' + model_path)
print('-' * 80)
y = y.astype(int)
if "test" not in model_path:
print('GroundTruth speech:\n' + '\n' + int_sequence_to_text(y))
else:
print('GroundTruth speech:\n' + '\n' +
int_sequence_to_text_test(y))
print('-' * 80)
print(pred_ints)
if "test" not in model_path:
print('Predicted speech:\n' + '\n' +
''.join(int_sequence_to_text(pred_ints)))
else:
print('Predicted speech:\n' + '\n' +
''.join(int_sequence_to_text_test(pred_ints)))
output_matrix(prediction[0])
def run_session(self, args):
# get data
################################
# get data and model handler
################################
# training data
data, label, _ = data_specification(args.mode, args.data_dir, 'npy')
# 暂时设置为test 数据变化之后会进行修改
dev_data, dev_label, _ = data_specification('test', args.data_dir,
'npy')
max_sequence_length = 0
batch_num = len(data) // args.batch_size
dev_batch_num = len(dev_data) // args.batch_size
for i in data:
max_sequence_length = max(max_sequence_length, i.shape[1])
for i in dev_data:
max_sequence_length = max(max_sequence_length, i.shape[1])
# Checkpointer
checkpointer = ModelCheckpoint(filepath=savedir + '/' +
args.model_name,
verbose=1)
# Training Phase
model = Deepspeech2(args, max_sequence_length)
history = model.model_2.fit_generator(
generator=create_batch(args, data, max_sequence_length, label),
validation_data=create_batch(args, dev_data, max_sequence_length,
dev_label),
steps_per_epoch=batch_num,
validation_steps=dev_batch_num,
epochs=args.epochs,
verbose=1,
callbacks=[checkpointer])
# save model loss
with open(savedir + '/' + args.pickle_name, 'wb') as f:
pickle.dump(history.history, f)
def get_data(mode='train',
data_dir='./PreferredStimuliData',
data_suffix='npy'):
"""Return data that is used for this visualization test
reuse the function inside the data helper, only return data with
batchsize with 1
"""
data, label, total_path_name = data_specification(mode, data_dir,
data_suffix)
max_sequence_length = 0
for i in data:
max_sequence_length = max(max_sequence_length, i.shape[1])
batch = create_batch(args, data, max_sequence_length, label)
inputs, outputs = next(batch)
#x = inputs['the_input']
#y = inputs['the_labels'][0]
return inputs, outputs, max_sequence_length