def train(self):
self.iterLineEdit.setText("xcz")
sample, label = get_train_data()
sample = np.array(sample, dtype='float')
sample = (sample) / 256.0
samp_num = len(sample)
inp_num = len(sample[0])
out_num = 10
hid_num = 15
loss = 0
self.BP = BPNetwork(inp_num, 20, hid_num, out_num, 0.1, 0.1, 0.1)
for step in range(0, 3):
if step == 1:
self.BP.reduce_lr(0.01, 0.01, 0.01)
elif step == 2:
self.BP.reduce_lr(0.001, 0.001, 0.001)
for i in range(0, samp_num):
train_label = np.zeros(out_num)
train_label[label[i]] = 1
self.BP.forward(sample[i])
self.BP.backward(np.array(train_label))
if i % 10000 == 0:
print(str(i + 60000 * step))
error = self.BP.erro
loss = 0
for j in range(0, len(error)):
loss = loss + abs(error[j])
print(loss)
self.iterLineEdit.setText("ending")
self.lossLineEdit.setText("ending")
def train():
start_time = time.time()
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
with tf.Graph().as_default():
with tf.name_scope('input'):
x = tf.placeholder(tf.float32, [None, IMAGE_SIZE * IMAGE_SIZE],
name='x-input')
y_ = tf.placeholder(tf.float32, [None, 4], name='y-input')
with tf.name_scope('input_reshape'):
image_shaped_input = tf.reshape(x, [-1, IMAGE_SIZE, IMAGE_SIZE, 1])
tf.summary.image('input', image_shaped_input, 10)
with tf.name_scope('dropout_keep_prob'):
keep_prob = tf.placeholder(tf.float32)
y = core.inference(image_shaped_input, keep_prob=keep_prob)
loss = core.loss(y, y_)
train_op = core.train(loss, FLAGS.learning_rate)
accuracy = core.evaluation(y, y_)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(init)
for i in range(20):
images, labels = input_data.get_train_data()
image_num, _ = images.shape
batch_num = int(image_num / 100)
random_index = random.sample(range(batch_num), batch_num)
for j in range(batch_num):
step = i * batch_num + j
index = random_index[j]
xs = images[index * 100:(index + 1) * 100]
ys = labels[index * 100:(index + 1) * 100]
if step % 50 == 0:
feed_dict = {x: xs, y_: ys, keep_prob: 1.0}
summary_str, acc = sess.run([summary, accuracy],
feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
print("step %d, train accuracy %g" % (step, acc))
else:
feed_dict = {
x: xs,
y_: ys,
keep_prob: FLAGS.dropout_keep_prob
}
summary_str, loss_value, _ = sess.run(
[summary, loss, train_op], feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
if (step + 1) % 1000 == 0:
saver.save(sess, checkpoint_file, global_step=step)
saver.save(sess, checkpoint_file)
duration = time.time() - start_time
print('%d seconds' % int(duration))
def run_eval(chkpt):
global net
net = imp.load_source('net', FLAGS.net_module)
with tf.Graph().as_default(), tf.device('/cpu:0'):
train_phase = tf.constant(False, name='train_phase', dtype=tf.bool)
t_images, t_labels = input_data.get_train_data(FLAGS.data_dir)
aux = {
'mean': np.mean(t_images, axis=0),
'std': np.std(t_images, axis=0)
}
v_images, v_labels = input_data.get_validation_data(FLAGS.data_dir)
images_ph = tf.placeholder(tf.float32,
shape=[None] + list(t_images.shape[1:]),
name='images_ph')
labels_ph = tf.placeholder(tf.int32, shape=[None], name='labels_ph')
images = net.aug_eval(images_ph, aux)
with tf.device('/gpu:0'):
with tf.name_scope('tower_0') as scope:
loss, evaluation = tower_loss_and_eval(images, labels_ph,
train_phase)
variable_averages = tf.train.ExponentialMovingAverage(0.999)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
saver = tf.train.Saver(tf.global_variables())
ema_saver = tf.train.Saver(variable_averages.variables_to_restore())
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
saver.restore(sess, chkpt)
ema_saver.restore(sess, chkpt)
sys.stdout.write('Checkpoint "%s" restored.\n' % (chkpt))
evaluate(sess, loss, evaluation, 'Train', images_ph, t_images,
labels_ph, t_labels)
evaluate(sess, loss, evaluation, 'Validation', images_ph, v_images,
labels_ph, v_labels)
def train(net, net_para, label, keep_prob, save_dir, log_dir):
times_1000 = net_para.train_steps/1000
summary_writer = tf.summary.FileWriter(log_dir)
for t in range(35,int(times_1000)):
tf.reset_default_graph()
graph = tf.Graph()
with graph.as_default() as g:
data_iterator = get_train_data(DATA_PATH, BATCH_SIZE)
next_element = data_iterator.get_next()
x = tf.placeholder(
tf.float32,
[BATCH_SIZE, net_para.image_size,
net_para.image_size, NUMBER_CHANNEL],
name = 'input-x'
)
y_ = tf.placeholder(
tf.int64,
[None],
name = 'input-y'
)
if NET_TYPE == 'mobilenet_0.5' or NET_TYPE == 'mobilenet_0.75':
model = net(x, label, keep_prob, net_para.skip, train_list=net_para.train_list, depth_multiplier=net_para.depth_multiplier)
else:
model = net(x, label, keep_prob, net_para.skip, train_list=net_para.train_list)
y = model.get_prediction()
global_step = tf.Variable(t*1000, trainable = False)
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
variable_averages_op = variable_averages.apply(tf.trainable_variables())
logits = y+1e10
cross_entropy_mean = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_))
with tf.name_scope('loss'):
loss = cross_entropy_mean
tf.summary.scalar('loss', loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(y,axis=1),y_)
correct_rate = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('accuracy', correct_rate)
if isinstance(model, MobileNets) and (TRAIN_MODEL == 'finetune' or TRAIN_MODEL == 'parttune'):
train_step = tf.train.RMSPropOptimizer(net_para.lr, net_para.lr_decay).minimize(loss, global_step=global_step,
var_list = tf.get_collection('train'))
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step,
# var_list = tf.get_collection('train'))
elif isinstance(model, MobileNets_depth) and (TRAIN_MODEL == 'finetune' or TRAIN_MODEL == 'parttune'):
train_step = tf.train.RMSPropOptimizer(net_para.lr, net_para.lr_decay).minimize(loss, global_step=global_step,
var_list = tf.get_collection('train'))
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step,
# var_list = tf.get_collection('train'))
else:
train_step = tf.train.RMSPropOptimizer(net_para.lr, net_para.lr_decay).minimize(loss, global_step=global_step)
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
with tf.control_dependencies([train_step, variable_averages_op]):
train_op = tf.no_op(name='train')
saver = tf.train.Saver()
merged = tf.summary.merge_all()
with tf.Session() as sess:
summary_writer.add_graph(sess.graph, t*1000)
if t == 0:
sess.run(tf.global_variables_initializer())
model.loadModel(sess)
else:
ckpt = tf.train.get_checkpoint_state(save_dir)
print('load model path is {0}'.format(ckpt.model_checkpoint_path))
saver.restore(sess, ckpt.model_checkpoint_path)
for i in range(1000):
xs, ys = next_element
ys = tf.reshape(ys,[BATCH_SIZE])
x_input, y_input = sess.run([xs,ys])
y_input -= 1
_, rate, loss_value, step, summary = sess.run([train_op, correct_rate, loss, global_step, merged], feed_dict={x: x_input, y_: y_input})
summary_writer.add_summary(summary,step)
if i%1000 == 0:
print("After {0:d} training step(s), loss on trian batch {1:g}".format(step, loss_value))
print("After {0:d} training step(s), correct rate on trian batch {1:s}".format(step, str(rate.astype(np.float))))
saver.save(sess, os.path.join(save_dir, MODEL_NAME), global_step=global_step)
summary_writer.close()
def train(net, net_para, label, keep_prob, save_dir, log_dir):
data_iterator = get_train_data(DATA_PATH, BATCH_SIZE)
next_element = data_iterator.get_next()
# x_mean = np.load('./Vehicle-Make-and-Model-CNN/data/'+MEAN_VALUE)
x = tf.placeholder(
tf.float32,
[BATCH_SIZE, net_para.image_size, net_para.image_size, NUMBER_CHANNEL],
name='input-x')
y_ = tf.placeholder(tf.int64, [None], name='input-y')
model = net(x,
label,
keep_prob,
net_para.skip,
train_list=net_para.train_list)
y = model.get_prediction()
global_step = tf.Variable(0, trainable=False)
variable_averages = tf.train.ExponentialMovingAverage(
MOVING_AVERAGE_DECAY, global_step)
variable_averages_op = variable_averages.apply(tf.trainable_variables())
# one_hot_y = tf.one_hot(y_, LABEL)
# cross_entropy = one_hot_y*tf.log(y+1e10)
# cross_entropy_mean = tf.reduce_mean(cross_entropy)
logits = y + 1e10
cross_entropy_mean = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=y_))
with tf.name_scope('loss'):
loss = cross_entropy_mean
tf.summary.scalar('loss', loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(y, axis=1), y_)
correct_rate = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('accuracy', correct_rate)
# learning_rate = tf.train.exponential_decay(
# net_para.lr,
# global_step,
# net_para.train_steps / BATCH_SIZE,
# net_para.lr_decay
# )
if isinstance(model, MobileNets) and (TRAIN_MODEL == 'finetune'
or TRAIN_MODEL == 'parttune'):
train_step = tf.train.RMSPropOptimizer(
net_para.lr,
net_para.lr_decay).minimize(loss,
global_step=global_step,
var_list=tf.get_collection('train'))
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step,
# var_list = tf.get_collection('train'))
else:
train_step = tf.train.RMSPropOptimizer(
net_para.lr, net_para.lr_decay).minimize(loss,
global_step=global_step)
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
with tf.control_dependencies([train_step, variable_averages_op]):
train_op = tf.no_op(name='train')
saver = tf.train.Saver()
merged = tf.summary.merge_all()
with tf.Session() as sess:
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
model.loadModel(sess)
for i in range(net_para.train_steps):
xs, ys = next_element
ys = tf.reshape(ys, [BATCH_SIZE])
x_input, y_input = sess.run([xs, ys])
# x_input -= x_mean
y_input -= 1
_, rate, loss_value, step, summary = sess.run(
[train_op, correct_rate, loss, global_step, merged],
feed_dict={
x: x_input,
y_: y_input
})
# print('-- origin --')
# print(y_input)
# print('-- predict --')
# print(np.argmax(y_pred,axis=1))
# print('-- compare--')
# print(np.equal(y_input, np.argmax(y_pred,axis=1)))
# print('-- rate --')
# pre_rate = np.sum(np.equal(y_input, np.argmax(y_pred,axis=1))) / BATCH_SIZE
# print(pre_rate)
# print('-- loss --')
# print('loss is {0}'.format(loss_value))
if i % 1000 == 0:
print(
"After {0:d} training step(s), loss on trian batch {1:g}".
format(step, loss_value))
print(
"After {0:d} training step(s), correct rate on trian batch {1:s}"
.format(step, str(rate.astype(np.float))))
saver.save(sess,
os.path.join(save_dir, MODEL_NAME),
global_step=global_step)
summary_writer.add_summary(summary, i)
summary_writer.close()
