def calc_motion_capa_data(file_name, file_dir, motion_out_dir, capa_out_dir): d = data_reader.Data() d.read(os.path.join(file_dir, file_name + ".txt")) print(u, d.task_id) t = get_vad_chunks(file_dir, file_name) # print(t) end = find_suitable_end(t, 0.4, 4.0) if calc_motion: out_file = os.path.join(motion_out_dir, d.task_id + '.txt') print(out_file) output = open(out_file, 'w', encoding='utf-8') output.write(d.user_pos + '\n') output.write(d.start_pos + '\n') output.write(d.description + '\n') output.write(d.hand + '\n') extract_motion_feature(0.05, end, d, output) output.close() if calc_capa: out_file = os.path.join(capa_out_dir, d.task_id + '.txt') output = open(out_file, 'w', encoding='utf-8') # feature = capa_feature.extract_time_feature(d, end * 1000, (end + 2.0) * 1000) feature = capa_feature.extract_time_feature_count_appearance_only(d, end * 1000, (end + 2.0) * 1000) for f in feature: output.write(str(f) + '\n') output.close()
def visualize_file(file_path):
data = data_reader.Data()
data.read(file_path)
type_list = data.get_types()
type_list.remove('TOUCH')
type_list.remove('CAPACITY')
type_list.sort()
n = 0
plt.figure(figsize=(10, 40))
for t in type_list:
n += 1
plt.subplot(len(type_list), 1, n)
plt.xlim(0, data.get_max_time())
frame_list = data.get_list(t)
plt.title(t)
for i in range(frame_list.get_data_way()):
if len(frame_list.value[i]) <= 100:
plot_format = 'x:'
else:
plot_format = '-'
plt.plot(frame_list.time_stamp, frame_list.value[i], plot_format, label=i)
plt.legend()
plt.suptitle(file_path, x=0.02, y=0.998, horizontalalignment='left')
plt.show()
def read_data(file_path_list): data_list = [] for file_path in file_path_list: d = data_reader.Data() d.read(file_path) data_list.append(d) return data_list
def calc_data(file_name, file_dir, out_dir):
print(file_name)
d = data_reader.Data()
d.read(os.path.join(file_dir, file_name + ".txt"))
out_file = os.path.join(out_dir, d.task_id + ".txt")
print(out_file)
output = open(out_file, 'w', encoding='utf-8')
output.write(d.user_pos + '\n')
output.write(d.start_pos + '\n')
output.write(d.description + '\n')
output.write(d.hand + '\n')
task = int(d.task_id.split("_")[0])
'''
# 1s
if task < 32 or d.description == '接听':
t = webrtcvad_utils.calc_vad(3, os.path.join(file_dir, file_name + ".wav"))
print(t)
if d.start_pos == '裤兜':
end = find_suitable_end(t, 4.0, 10.0)
else:
end = find_suitable_end(t, 1.0, 4.0)
extract_feature(end - 1.0, end, d, output)
else:
max_time = d.get_max_time() / 1000
start = 2.0
while start + 2.5 < max_time:
extract_feature(start, start + 1.0, d, output)
start += 2.0
'''
# 2s
if task < 32 or d.description == '接听':
t = webrtcvad_utils.calc_vad(
3, os.path.join(file_dir, file_name + ".wav"))
print(t)
if d.start_pos == '裤兜':
end = find_suitable_end(t, 5.0, 10.0)
else:
end = find_suitable_end(t, 2.0, 4.0)
extract_feature(end - 2.0, end, d, output)
else:
max_time = d.get_max_time() / 1000
start = 1.0
while start + 2.5 < max_time:
extract_feature(start, start + 2.0, d, output)
start += 2.0
def calc_motion_capa_data(file_name, file_dir):
d = data_reader.Data()
d.read(os.path.join(file_dir, file_name + '.txt'))
print(u, d.task_id)
t = get_vad_chunks(file_dir, file_name)
end, voice_end = find_suitable_end(t, 0.4, 4.0)
if calc_motion:
for t in time_point:
extract_motion_feature(0.05, end, voice_end, d, t)
if calc_capa:
for t in time_point:
extract_capa_feature(0.05, end, voice_end, d, t)
def train():
'''kernel train
'''
with tf.Session(config=config) as sess:
saver = tf.train.Saver()
global_step = 0
iter = 1
current_learning_rate = cfg.init_learning_rate
init_variable = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_variable)
data = dtrd.Data()
data.load(data_filename=cfg.train_dataset)
puts_debug('data size: {}'.format(data.size()))
while global_step < data.size() * cfg.epoch:
batch_x, _, _, batch_y = data.decode_and_fetch(
batch_size=cfg.batch_size)
_, loss_val = sess.run([train_op, loss_op],
feed_dict={
xs: batch_x,
ys: batch_y,
learning_rate: current_learning_rate
})
puts_info('iter: {}, loss: {}'.format(iter, loss_val))
global_step += cfg.batch_size
if iter % cfg.test_iter == 0:
batch_x, _, _, batch_y = data.decode_and_fetch(
batch_size=cfg.batch_size)
accuracy_val, loss_val = sess.run([accuracy_op, loss_op],
feed_dict={
xs:
batch_x,
ys:
batch_y,
learning_rate:
current_learning_rate
})
puts_info('accuracy: {:.4f}, loss: {:.4f}'.format(
accuracy_val, loss_val))
if iter % cfg.save_iter == 0:
saver.save(
sess,
os.path.join(cfg.save_path,
'model.ckpt-' + str(global_step)))
puts_info(
'iter: {}, model has been saved under {}/model.ckpt-{}'.
format(iter, cfg.save_path, global_step))
iter += 1
batch_x, _, _, batch_y = data.decode_and_fetch(
batch_size=cfg.batch_size)
accuracy_val, loss_val, outputs_val = sess.run(
[accuracy_op, loss_op, outputs_op],
feed_dict={
xs: batch_x,
ys: batch_y,
learning_rate: current_learning_rate
})
puts_info('final >> val: \n{}, accuracy: {:.4f}, loss: {:.4f}'.format(
outputs_val, accuracy_val, loss_val))
saver.save(sess, os.path.join(cfg.save_path, 'model.ckpt'))
puts_info('final model has been saved under {}'.format(
os.path.join(cfg.save_path, 'model.ckpt')))