def train_act(features_train, labels_train, features_test, labels_test):
sess = tf.InteractiveSession()
x = tf.placeholder(
dtype=tf.float32,
shape=[None, FLAGS.IMG_WIDTH, FLAGS.IMG_HEIGHT, FLAGS.IMG_CHANNEL],
name='features')
y = tf.placeholder(dtype=tf.float32,
shape=[None, FLAGS.cls],
name='labels')
pred = model(data=x)
with tf.name_scope('loss'):
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
tf.summary.scalar('loss', loss)
tf.summary.histogram('loss', loss)
with tf.name_scope('train'):
train = tf.train.MomentumOptimizer(
learning_rate=FLAGS.LEARNING_RATE,
momentum=FLAGS.MOMENTUM).minimize(loss)
with tf.name_scope('accuracy'):
accuracy = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1)), tf.float32))
tf.summary.scalar('accuracy', accuracy)
tf.summary.histogram('accuracy', accuracy)
merge = tf.summary.merge_all()
logwriter = tf.summary.FileWriter(FLAGS.LOG_DIR, sess.graph)
initial = tf.global_variables_initializer()
sess.run(initial)
data_size = features_train.shape[0]
iterations = int(data_size / FLAGS.batch_size)
for _ in range(FLAGS.epoches):
for i in range(iterations):
data = []
labels = []
if i == iterations - 1:
data = features_train[i * FLAGS.batch_size:data_size, :, :, :]
labels = labels_train[i * FLAGS.batch_size:data_size]
else:
data = features_train[i * FLAGS.batch_size:(i + 1) *
FLAGS.batch_size, :, :, :]
labels = labels_train[i * FLAGS.batch_size:(i + 1) *
FLAGS.batch_size]
sess.run(train, feed_dict={x: data, y: labels})
if i % 10 == 0:
summary, accuracy_res = sess.run([merge, accuracy],
feed_dict={
x: features_test,
y: labels_test
})
logwriter.add_summary(summary, i)
print(visualize.get_time() +
' epoch %d, train_iteration at %d, test score: %f ' %
(_, i, accuracy_res))
sess.close()
logwriter.close()
import tensorflow as tf
from tools import visualize, dataset
NUMBER_CLASSES = 10
IMAGE_WIDTH = 32
IMAGE_HEIGHT = 32
IMAGE_CHANNELS = 3
DATASET_PATH = '/home/workspace/cifar10/pythonver'
LEARNING_RATE = 0.001
MOMENTUM = 0.9
batch_size = 128
epoches = 5
KEEP_PROB = 0.6
LOG_DIR = 'Log--' + visualize.get_time()
def convolution(data, kernel, strides, name, bias, padding='SAME'):
with tf.name_scope(name):
conv = tf.nn.conv2d(input=data,
filter=kernel,
strides=[1, strides, strides, 1],
padding=padding,
name=name)
with tf.name_scope(name + '_relu'):
conv = tf.nn.bias_add(conv, bias)
relu = tf.nn.relu(conv, name=name + '_relu')
return relu
return conv
tf.app.flags.DEFINE_boolean('train_mode', True, 'True--train, False--eval')
tf.app.flags.DEFINE_boolean('bottleneck_residual_flag', True,
'True--bottleneck residual, False--residual')
tf.app.flags.DEFINE_integer('residual_numbers', 50,
'numbers of residual module')
tf.app.flags.DEFINE_integer('cls', 10, 'numbers of classifiers')
tf.app.flags.DEFINE_integer('IMG_WIDTH', 32, 'image width')
tf.app.flags.DEFINE_integer('IMG_HEIGHT', 32, 'image height')
tf.app.flags.DEFINE_integer('IMG_CHANNEL', 3, 'image channel')
tf.app.flags.DEFINE_integer('batch_size', 12, 'batch size')
tf.app.flags.DEFINE_integer('epoches', 1, 'epoches')
tf.app.flags.DEFINE_float('LEARNING_RATE', 0.1,
'learning rate for momentum GD')
tf.app.flags.DEFINE_float('MOMENTUM', 0.9, 'momentum rate for momentum GD')
tf.app.flags.DEFINE_string('data_path', 'cifar10-pythonver', 'path of dataset')
tf.app.flags.DEFINE_string('LOG_DIR', 'Log--' + visualize.get_time(),
'tensorboard log dir')
def convolution(data, kernel, strides, name, padding='SAME'):
with tf.name_scope(name):
conv = tf.nn.conv2d(input=data,
filter=kernel,
strides=[1, strides, strides, 1],
padding=padding,
name=name)
return conv
def pooling(data, ksize, strides, name, padding='VALID'):
with tf.name_scope(name):
import tensorflow as tf
import six
from tools import visualize
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_integer('cls', 10, 'numbers of classifiers')
tf.app.flags.DEFINE_integer('IMG_WIDTH', 299, 'image width')
tf.app.flags.DEFINE_integer('IMG_HEIGHT', 299, 'image height')
tf.app.flags.DEFINE_integer('IMG_CHANNEL', 3, 'image channel')
tf.app.flags.DEFINE_integer('batch_size', 5, 'batch size')
tf.app.flags.DEFINE_integer('epoches', 1, 'epoches')
tf.app.flags.DEFINE_float('LEARNING_RATE', 0.0001, 'learning rate for momentum GD')
tf.app.flags.DEFINE_float('MOMENTUM', 0.9, 'momentum rate for momentum GD')
tf.app.flags.DEFINE_string('data_path', '/home/workspace/cc', 'path of dataset')
tf.app.flags.DEFINE_string('LOG_DIR', 'Log--' + visualize.get_time(), 'tensorboard log dir')
def convolution(data, kernel, strides, padding='SAME', name=None):
with tf.name_scope(name):
conv = tf.nn.conv2d(input=data,
filter=tf.Variable(tf.truncated_normal(shape=kernel, stddev=1e-2, dtype=tf.float32)),
strides=[1, strides, strides, 1], padding=padding, name=name)
return conv
def pooling(data, ksize, strides, padding='SAME', type='max', name=None):
with tf.name_scope(name):
if type == 'max':
pool = tf.nn.max_pool(value=data, ksize=[1, ksize, ksize, 1],
strides=[1, strides, strides, 1], padding=padding, name=name)