def load_images(self):
"""
load visualization samples.
:return
next_element_train (generator): can iter images and labels.
"""
loader = LoadData(self.batch_size, self.sess)
next_element_train = loader.get_data(self.train_files)
return next_element_train
def Caculate(self,
sess,
files,
accuracy_tensor,
Cost_tensor,
data_placeholder,
labels_placeholder,
rate_placeholder,
model='train'):
"""
Calculate accuracy and loss.
:param sess : tensor graph.
:param files (ndarray): score data.
:param accuracy_tensor (tensor): accuracy tensor function graph.
:param Cost_tensor (tensor): cost tensor function graph.
:param data_placeholder (tensor): data placeholder.
:param labels_placeholder (tensor): labels placeholder.
:param rate_placeholder (tensor): rate placeholder.
:param model (string): can choose 'train' or 'test' to scored model.
:return acc_mean,loss_mean (float): mean accuracy and mean loss.
"""
if model == 'train':
N = self.N_train
else:
N = self.N_test
loader = LoadData(self.batch_size, sess)
next_element_ = loader.get_data(files)
acc, loss, count = 0, 0, 1
while 1:
try:
images, labels = sess.run(next_element_)
print('Score {}/{} \r'.format(count, N), end='', flush=True)
count += 1
except tf.errors.OutOfRangeError:
break
else:
acc_, loss_ = sess.run(
[accuracy_tensor, Cost_tensor],
feed_dict={
data_placeholder: images,
labels_placeholder: labels,
rate_placeholder: 0
})
acc += acc_
loss += loss_
acc_mean = acc / count
loss_mean = loss / count
return acc_mean, loss_mean
def fit(self, lr, epochs, drop_rate):
"""
fitting model.
:param lr (float): learning rate.
:param epochs (int): Iterate of epoch.
:param drop_rate (float): dropout rate. rate = 1 - keep_prob.
:return:
"""
# create placeholder
data = tf.placeholder(tf.float32, [None, 227, 227, 3], name='Input')
labels = tf.placeholder(tf.float32, [None, 1], name='Labels')
rate = tf.placeholder(tf.float32, name='rate')
# build model.
params = self.init_parameters()
out = self.forward(data, params, rate)
Cost = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=out, labels=labels))
optimizer = tf.train.RMSPropOptimizer(learning_rate=lr).minimize(Cost)
# score.
predict = tf.round(tf.sigmoid(out))
equal = tf.equal(labels, predict)
correct = tf.cast(equal, tf.float32)
accuracy = tf.reduce_mean(correct)
# split date..
split_data = SplitData(self.file_dir,
Load_samples=self.Load_samples,
test_rate=self.test_rate)
train_files, test_files = split_data()
self.N_train = len(train_files) // self.batch_size
self.N_test = len(test_files) // self.batch_size
# Saver..
saver = tf.train.Saver()
tf.add_to_collection('pre_network', out)
init = tf.global_variables_initializer()
# training...
with tf.Session() as sess:
sess.run(init)
for epoch in range(epochs):
loader = LoadData(self.batch_size, sess)
next_element_train = loader.get_data(train_files)
# running all training set...
count = 1
while 1:
try:
images, target = sess.run(next_element_train)
print('Training {}/{} \r'.format(count, self.N_train),
end='',
flush=True)
count += 1
except tf.errors.OutOfRangeError:
break
else:
_ = sess.run(optimizer,
feed_dict={
data: images,
labels: target,
rate: drop_rate
})
acc_train, loss_train = self.Caculate(sess, train_files,
accuracy, Cost, data,
labels, rate, 'train')
acc_test, loss_test = self.Caculate(sess, test_files, accuracy,
Cost, data, labels, rate,
'test')
print(
'[{}/{}] train loss:{:.4f} - train acc:{:.4f} - test loss:{:.4f} - test acc:{:.4f}'
.format(epoch + 1, epochs, loss_train, acc_train,
loss_test, acc_test))
if acc_train >= 0.980:
break
# Saver ....
saver.save(sess, 'model/alexNet')
