x_train = util.pre_treat(x_train_orig)
x_test = util.pre_treat(x_test_orig)
y_train = util.pre_treat(y_train_orig, is_x=False, class_num=len(classes))
y_test = util.pre_treat(y_test_orig, is_x=False, class_num=len(classes))
cnn = Cnn(config.conv_layers, config.fc_layers, config.filters,
config.learning_rate, config.beta1, config.beta2)
(m, n_H0, n_W0, n_C0) = x_train.shape
n_y = y_train.shape[1]
# construction calculation graph
cnn.initialize(n_H0, n_W0, n_C0, n_y)
cnn.forward()
cost = cnn.cost()
optimizer = cnn.get_optimizer(cost)
predict, accuracy = cnn.predict()
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
for i in range(1, config.num_epochs + 1):
num_mini_batches = int(m / config.mini_batch_size)
# seed += 1
mini_batches = util.random_mini_batches(x_train, y_train,
config.mini_batch_size)
cost_value = 0
for mini_batch in mini_batches:
def train():
TIMESTAMP = "{0:%Y-%m-%d-%H-%M/}".format(datetime.now())
log.log_info('program start')
data, num_good, num_bad = util.load_train_data(num_data // 2)
log.log_debug('Data loading completed')
# resample
data, length = util.resample(data, 600)
data = util.reshape(data, length)
good_data_origin = data[:num_good, :]
bad_data_origin = data[num_good:, :]
# extract bad data for test and train
permutation = list(np.random.permutation(len(bad_data_origin)))
shuffled_bad_data = bad_data_origin[permutation, :]
test_bad_data = shuffled_bad_data[:int(num_bad * 0.3), :]
train_bad_data_origin = shuffled_bad_data[int(num_bad * 0.3):, :]
# extract corresponding good data for test and train
permutation = list(np.random.permutation(len(good_data_origin)))
shuffled_good_data = good_data_origin[permutation, :]
test_good_data = shuffled_good_data[:len(test_bad_data), :]
train_good_data = shuffled_good_data[len(test_bad_data):, :]
assert len(test_bad_data) == len(test_good_data)
# construct test data
test_y = np.array([1.] * len(test_good_data) + [0.] * len(test_bad_data), dtype=np.float).reshape(
(len(test_bad_data) + len(test_good_data), 1))
test_x = np.vstack((test_good_data, test_bad_data))
# expand the number of bad data for train
train_x = np.vstack((train_good_data, train_bad_data_origin))
train_y = np.array([1.] * len(train_good_data) + [0.] * len(train_bad_data_origin), dtype=np.float).reshape(
(len(train_bad_data_origin) + len(train_good_data), 1))
train_x, train_y, num_expand = util.expand(train_x, train_y)
# regularize
for i in range(len(train_x)):
train_x[i, :, 0] = util.regularize(train_x[i, :, 0])
train_x[i, :, 1] = util.regularize(train_x[i, :, 1])
train_x[i, :, 2] = util.regularize(train_x[i, :, 2])
for i in range(len(test_x)):
test_x[i, :, 0] = util.regularize(test_x[i, :, 0])
test_x[i, :, 1] = util.regularize(test_x[i, :, 1])
test_x[i, :, 2] = util.regularize(test_x[i, :, 2])
# random
train_x, train_y = util.shuffle_data(train_x, train_y)
log.log_debug('prepare completed')
log.log_info('convolution layers: ' + str(conv_layers))
log.log_info('filters: ' + str(filters))
log.log_info('full connected layers: ' + str(fc_layers))
log.log_info('learning rate: %f' % learning_rate)
log.log_info('keep prob: ' + str(keep_prob))
log.log_info('the number of expanding bad data: ' + str(num_expand))
log.log_info('mini batch size: ' + str(mini_batch_size))
if mini_batch_size != 0:
assert mini_batch_size <= len(train_x)
cnn = Cnn(conv_layers, fc_layers, filters, learning_rate)
(m, n_W0, n_C0) = train_x.shape
n_y = train_y.shape[1]
# construction calculation graph
cnn.initialize(n_W0, n_C0, n_y)
cost = cnn.cost()
optimizer = cnn.get_optimizer(cost)
predict, accuracy = cnn.predict()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
# log for tensorboard
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter("resource/tsb/train/" + TIMESTAMP, sess.graph)
test_writer = tf.summary.FileWriter("resource/tsb/test/" + TIMESTAMP)
if enable_debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
sess.run(init)
for i in range(1, num_epochs + 1):
if mini_batch_size != 0:
num_mini_batches = int(m / mini_batch_size)
mini_batches = util.random_mini_batches(train_x, train_y, mini_batch_size)
cost_value = 0
for mini_batch in mini_batches:
(mini_batch_x, mini_batch_y) = mini_batch
_, temp_cost = sess.run([optimizer, cost], feed_dict={cnn.x: mini_batch_x, cnn.y: mini_batch_y,
cnn.keep_prob: keep_prob})
cost_value += temp_cost
cost_value /= num_mini_batches
else:
_, cost_value = sess.run([optimizer, cost],
feed_dict={cnn.x: train_x, cnn.y: train_y, cnn.keep_prob: keep_prob})
# disable dropout
summary_train, train_accuracy = sess.run([merged, accuracy],
feed_dict={cnn.x: train_x, cnn.y: train_y,
cnn.keep_prob: 1})
summary_test, test_accuracy = sess.run([merged, accuracy],
feed_dict={cnn.x: test_x, cnn.y: test_y, cnn.keep_prob: 1})
train_writer.add_summary(summary_train, i - 1)
test_writer.add_summary(summary_test, i - 1)
if print_detail and (i % 10 == 0 or i == 1):
info = '\nIteration %d\n' % i + \
'Cost: %f\n' % cost_value + \
'Train accuracy: %f\n' % train_accuracy + \
'Test accuracy: %f' % test_accuracy
log.log_info(info)
# stop when test>0.95 and train>0.99
if test_accuracy >= 0.95 and train_accuracy >= 0.99:
info = '\nIteration %d\n' % i + \
'Cost: %f\n' % cost_value + \
'Train accuracy: %f\n' % train_accuracy + \
'Test accuracy: %f' % test_accuracy
log.log_info(info)
saver.save(sess, "resource/model/" + TIMESTAMP)
break
saver.save(sess, "resource/model/" + TIMESTAMP)
train_writer.close()
test_writer.close()
log.log_info('program end')