Tensorflow:VGG-net训练过程总结
前言：原本我是比较偏向数据挖掘方法的，后来找到的工作更偏向CV和NLP，所以也在慢慢的往CV和NLP靠，有什么说的不好，望指正。
1、VGG-net简介
VGG-net是牛津大学和Deepmind一起研发的一种深度卷积神经网络，在2014年的ILSVRC比赛在取得第二名的成绩，将Top-5的错误率将到7.3%，而当年排名第一的正是大名鼎鼎的Google-Net，后面有空在详细说。VGG-net的泛化能力非常好，在许多不同的数据集都有较好的表现。2019年的现在，VGG仍然经常被用来提取特征图像，VGG对于新入门的同学也是比较友好（没有太多复杂的结构）。
2、问题解答
（1）什么是感受野？
在卷积神经网络中，感受野（Receptive Field）的定义是卷积神经网络每一层输出的特征图（feature map）上的像素点在输入图片上映射的区域大小。再通俗点的解释是，特征图上的一个点对应输入图上的区域，如下图所示。
感受野
2）为什么说连续两个3*3的卷积核等价于一个5*5的卷积核，连续三个3*3的卷积核等价于一个7*7的卷积核？
首先解释为什么要使用更小的卷积核，每一张图片使用的卷积神经网络的核是共享的，使用更小的核可以降低模型的参数量，然而较小的核函数的感受野较小，为了使得3*3核的感受野和5*5的感受野一致，VGG提出了使用连续的两个3*3卷积核来替代5*5的卷积核，参数量从5*5=25，降为3*3*2=18（两个3*3的卷积核参数量）。网上我也找到相关的问题（1）的解释，没看太懂，下面说下我的理解：
padding为VALID，假设一幅图片是w*w(不是一般性，设为32*32)，卷积核大小设为k*k，步长设为S，padding的像素个数P（一般情况为0），输出的图片尺寸为N*N，其中N=((w-k+2P)/S)+1。
1）对于一个5*5的卷积核，步长S为1，P=0，卷积后的输出尺寸边长为((32-5+2*0))/1+1=28,图片为28*28。
2）对于一个3*3的卷积核，其它设置一样，输出的尺寸为30*30，对得到的30*30的图片继续使用3*3卷积核，得到的输出的图片尺寸为28*28。
因此，连续使用两次3*3的卷积核得到的图片尺寸和一次使用5*5卷积核的尺寸一样，但参数量更少。
3、VGG-net的网络结构
VGG16经典模型图
VGG参数量计算（参数量和卷积核大小和个数有关）
4、训练过程
（1）导入相关包，设置模型参数
# -*- coding:utf-8 -*-
import tensorflow as tf
import numpy as np
import time
import os
import sys
import random
import pickle
import tarfile
class_num = 10
image_size = 32
img_channels = 3
batch_size = 250
iterations = 200 # 训练集50000=250*200
total_epoch = 200
weight_decay = 0.0003
dropout_rate = 0.5
momentum_rate = 0.9
log_save_path = './vgg_16_logs'
model_save_path = './model/'
（2）下载数据集
def download_data():
dirname = 'CAFIR-10_data/cifar-10-python' # 解压后的文件夹
origin = 'http://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz'
fname = './CAFIR-10_data/cifar-10-python.tar.gz'
#fname = './CAFIR-10_data'
fpath = './' + dirname
download = False
if os.path.exists(fpath) or os.path.isfile(fname):
download = False
print("DataSet already exist!")
import tarfile
if fname.endswith("tar.gz"):
tar = tarfile.open(fname, "r:gz")
tar.extractall()
tar.close()
#tarfile.open(fpath, "r:gz").extractall(fname)
elif fname.endswith("tar"):
tar = tarfile.open(fname, "r:")
tar.extractall()
tar.close()
else:
download = True
if download:
print('Downloading data from', origin)
import urllib.request
import tarfile
def reporthook(count, block_size, total_size):
global start_time
if count == 0:
start_time = time.time()
return
duration = time.time() - start_time
progress_size = int(count * block_size)
speed = int(progress_size / (1024 * duration))
percent = min(int(count*block_size*100/total_size),100)
sys.stdout.write("\r...%d%%, %d MB, %d KB/s, %d seconds passed" %
(percent, progress_size / (1024 * 1024), speed, duration))
sys.stdout.flush()
urllib.request.urlretrieve(origin, fname, reporthook)
print('Download finished. Start extract!', origin)
if fname.endswith("tar.gz"):
tar = tarfile.open(fname, "r:gz")
tar.extractall()
tar.close()
elif fname.endswith("tar"):
tar = tarfile.open(fname, "r:")
tar.extractall()
tar.close()
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='bytes')
return dict
def load_data_one(file):
batch = unpickle(file)
data = batch[b'data']
labels = batch[b'labels']
print("Loading %s : %d." % (file, len(data)))
return data, labels
def load_data(files, data_dir, label_count):
global image_size, img_channels
data, labels = load_data_one(data_dir + '/' + files[0])
for f in files[1:]:
data_n, labels_n = load_data_one(data_dir + '/' + f)
data = np.append(data, data_n, axis=0)
labels = np.append(labels, labels_n, axis=0)
labels = np.array([[float(i == label) for i in range(label_count)] for label in labels])
data = data.reshape([-1, img_channels, image_size, image_size])
data = data.transpose([0, 2, 3, 1])
return data, labels
def prepare_data():
print("======Loading data======")
download_data()
#data_dir = './CAFIR-10_data/cifar-10-python/cifar-10-batches-py'
data_dir = './cifar-10-batches-py' # 数据集解压后的文件夹
image_dim = image_size * image_size * img_channels
meta = unpickle(data_dir + '/batches.meta')
label_names = meta[b'label_names']
label_count = len(label_names)
train_files = ['data_batch_%d' % d for d in range(1, 6)]
train_data, train_labels = load_data(train_files, data_dir, label_count)
test_data, test_labels = load_data(['test_batch'], data_dir, label_count)
print("Train data:", np.shape(train_data), np.shape(train_labels))
print("Test data :", np.shape(test_data), np.shape(test_labels))
print("======Load finished======")
print("======Shuffling data======")
indices = np.random.permutation(len(train_data))
train_data = train_data[indices]
train_labels = train_labels[indices]
print("======Prepare Finished======")
return train_data, train_labels, test_data, test_labels
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape, dtype=tf.float32)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool(input, k_size=1, stride=1, name=None):
return tf.nn.max_pool(input, ksize=[1, k_size, k_size, 1], strides=[1, stride, stride, 1],
padding='SAME', name=name)
def batch_norm(input):
return tf.contrib.layers.batch_norm(input, decay=0.9, center=True, scale=True, epsilon=1e-3,
is_training=train_flag, updates_collections=None)
# random crop 随机裁剪,下面的数据增强
def _random_crop(batch, crop_shape, padding=None):
oshape = np.shape(batch[0])
if padding:
oshape = (oshape[0] + 2*padding, oshape[1] + 2*padding)
new_batch = []
npad = ((padding, padding), (padding, padding), (0, 0))
for i in range(len(batch)):
new_batch.append(batch[i])
if padding:
new_batch[i] = np.lib.pad(batch[i], pad_width=npad,
mode='constant', constant_values=0)
nh = random.randint(0, oshape[0] - crop_shape[0])
nw = random.randint(0, oshape[1] - crop_shape[1])
new_batch[i] = new_batch[i][nh:nh + crop_shape[0],
nw:nw + crop_shape[1]]
return new_batch
def _random_flip_leftright(batch):
for i in range(len(batch)):
if bool(random.getrandbits(1)):
batch[i] = np.fliplr(batch[i])
return batch
def data_augmentation(batch):
batch = _random_flip_leftright(batch)
batch = _random_crop(batch, [32, 32], 4)
return batchdef data_preprocessing(x_train,x_test):
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train[:, :, :, 0] = (x_train[:, :, :, 0] - np.mean(x_train[:, :, :, 0])) / np.std(x_train[:, :, :, 0])
x_train[:, :, :, 1] = (x_train[:, :, :, 1] - np.mean(x_train[:, :, :, 1])) / np.std(x_train[:, :, :, 1])
x_train[:, :, :, 2] = (x_train[:, :, :, 2] - np.mean(x_train[:, :, :, 2])) / np.std(x_train[:, :, :, 2])
x_test[:, :, :, 0] = (x_test[:, :, :, 0] - np.mean(x_test[:, :, :, 0])) / np.std(x_test[:, :, :, 0])
x_test[:, :, :, 1] = (x_test[:, :, :, 1] - np.mean(x_test[:, :, :, 1])) / np.std(x_test[:, :, :, 1])
x_test[:, :, :, 2] = (x_test[:, :, :, 2] - np.mean(x_test[:, :, :, 2])) / np.std(x_test[:, :, :, 2])
return x_train, x_test
def learning_rate_schedule(epoch_num):
if epoch_num < 81:
return 0.01#0.1
elif epoch_num < 121:
return 0.001#0.01
else:
def run_testing(sess, ep):
acc = 0.0
loss = 0.0
pre_index = 0
add = 1000
for it in range(10):
batch_x = test_x[pre_index:pre_index+add]
batch_y = test_y[pre_index:pre_index+add]
pre_index = pre_index + add
loss_, acc_ = sess.run([cross_entropy, accuracy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: False})
loss += loss_ / 10.0
acc += acc_ / 10.0
summary = tf.Summary(value=[tf.Summary.Value(tag="test_loss", simple_value=loss),
tf.Summary.Value(tag="test_accuracy", simple_value=acc)])
return acc, loss, summary
if __name__ == '__main__':
train_x, train_y, test_x, test_y = prepare_data()
train_x, test_x = data_preprocessing(train_x, test_x)
# define placeholder x, y_ , keep_prob, learning_rate
x = tf.placeholder(tf.float32,[None, image_size, image_size, 3])
y_ = tf.placeholder(tf.float32, [None, class_num])
keep_prob = tf.placeholder(tf.float32)
learning_rate = tf.placeholder(tf.float32)
train_flag = tf.placeholder(tf.bool)
# build_network
W_conv1_1 = tf.get_variable('conv1_1', shape=[3, 3, 3, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_1 = bias_variable([64])
output = tf.nn.relu(batch_norm(conv2d(x, W_conv1_1) + b_conv1_1))
W_conv1_2 = tf.get_variable('conv1_2', shape=[3, 3, 64, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_2 = bias_variable([64])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv1_2) + b_conv1_2))
output = max_pool(output, 2, 2, "pool1")
W_conv2_1 = tf.get_variable('conv2_1', shape=[3, 3, 64, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_1 = bias_variable([128])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv2_1) + b_conv2_1))
W_conv2_2 = tf.get_variable('conv2_2', shape=[3, 3, 128, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_2 = bias_variable([128])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv2_2) + b_conv2_2))
output = max_pool(output, 2, 2, "pool2")
W_conv3_1 = tf.get_variable('conv3_1', shape=[3, 3, 128, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_1 = bias_variable([256])
output = tf.nn.relu( batch_norm(conv2d(output,W_conv3_1) + b_conv3_1))
W_conv3_2 = tf.get_variable('conv3_2', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_2 = bias_variable([256])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv3_2) + b_conv3_2))
W_conv3_3 = tf.get_variable('conv3_3', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_3 = bias_variable([256])
output = tf.nn.relu( batch_norm(conv2d(output, W_conv3_3) + b_conv3_3))
output = max_pool(output, 2, 2, "pool3")
W_conv4_1 = tf.get_variable('conv4_1', shape=[3, 3, 256, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_1 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_1) + b_conv4_1))
W_conv4_2 = tf.get_variable('conv4_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_2 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_2) + b_conv4_2))
W_conv4_3 = tf.get_variable('conv4_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_3 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_3) + b_conv4_3))
output = max_pool(output, 2, 2)
W_conv5_1 = tf.get_variable('conv5_1', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_1 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_1) + b_conv5_1))
W_conv5_2 = tf.get_variable('conv5_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_2 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_2) + b_conv5_2))
W_conv5_3 = tf.get_variable('conv5_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_3 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_3) + b_conv5_3))
#output = max_pool(output, 2, 2)
# output = tf.contrib.layers.flatten(output)
output = tf.reshape(output, [-1, 2*2*512])
W_fc1 = tf.get_variable('fc1', shape=[2048, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc1 = bias_variable([4096])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc1) + b_fc1) )
output = tf.nn.dropout(output, keep_prob)
W_fc2 = tf.get_variable('fc7', shape=[4096, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc2 = bias_variable([4096])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc2) + b_fc2))
output = tf.nn.dropout(output, keep_prob)
W_fc3 = tf.get_variable('fc3', shape=[4096, 10], initializer=tf.contrib.keras.initializers.he_normal())
b_fc3 = bias_variable([10])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc3) + b_fc3))
# output = tf.reshape(output,[-1,10])
# loss function: cross_entropy
# train_step: training operation
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=output))
l2 = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
train_step = tf.train.MomentumOptimizer(learning_rate, momentum_rate, use_nesterov=True).\
minimize(cross_entropy + l2 * weight_decay)
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# initial an saver to save model
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(log_save_path,sess.graph)
# epoch = 164
# make sure [bath_size * iteration = data_set_number]
for ep in range(1, total_epoch+1):
lr = learning_rate_schedule(ep)
pre_index = 0
train_acc = 0.0
train_loss = 0.0
start_time = time.time()
print("\n epoch %d/%d:" % (ep, total_epoch))
for it in range(1, iterations+1):
batch_x = train_x[pre_index:pre_index+batch_size]
batch_y = train_y[pre_index:pre_index+batch_size]
batch_x = data_augmentation(batch_x)
_, batch_loss = sess.run([train_step, cross_entropy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: dropout_rate,
learning_rate: lr, train_flag: True})
batch_acc = accuracy.eval(feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: True})
train_loss += batch_loss
train_acc += batch_acc
pre_index += batch_size
if it == iterations:
train_loss /= iterations
train_acc /= iterations
loss_, acc_ = sess.run([cross_entropy, accuracy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: True})
train_summary = tf.Summary(value=[tf.Summary.Value(tag="train_loss", simple_value=train_loss),
tf.Summary.Value(tag="train_accuracy", simple_value=train_acc)])
val_acc, val_loss, test_summary = run_testing(sess, ep)
summary_writer.add_summary(train_summary, ep)
summary_writer.add_summary(test_summary, ep)
summary_writer.flush()
print("iteration: %d/%d, cost_time: %ds, train_loss: %.4f, "
"train_acc: %.4f, test_loss: %.4f, test_acc: %.4f"
% (it, iterations, int(time.time()-start_time), train_loss, train_acc, val_loss, val_acc))
else:
print("iteration: %d/%d, train_loss: %.4f, train_acc: %.4f"
% (it, iterations, train_loss / it, train_acc / it))
save_path = saver.save(sess, model_save_path)
print("Model saved in file: %s" % save_path)
代码汇总：
# -*- coding:utf-8 -*-
import tensorflow as tf
import numpy as np
import time
import os
import sys
import random
import pickle
import tarfile
class_num = 10
image_size = 32
img_channels = 3
batch_size = 250
iterations = 200 # 训练集50000=250*200
total_epoch = 200
weight_decay = 0.0003
dropout_rate = 0.5
momentum_rate = 0.9
log_save_path = './vgg_16_logs'
model_save_path = './model/'
def download_data():
dirname = 'CAFIR-10_data/cifar-10-python' # 解压后的文件夹
origin = 'http://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz'
fname = './CAFIR-10_data/cifar-10-python.tar.gz'
#fname = './CAFIR-10_data'
fpath = './' + dirname
download = False
if os.path.exists(fpath) or os.path.isfile(fname):
download = False
print("DataSet already exist!")
import tarfile
if fname.endswith("tar.gz"):
tar = tarfile.open(fname, "r:gz")
tar.extractall()
tar.close()
#tarfile.open(fpath, "r:gz").extractall(fname)
elif fname.endswith("tar"):
tar = tarfile.open(fname, "r:")
tar.extractall()
tar.close()
else:
download = True
if download:
print('Downloading data from', origin)
import urllib.request
import tarfile
def reporthook(count, block_size, total_size):
global start_time
if count == 0:
start_time = time.time()
return
duration = time.time() - start_time
progress_size = int(count * block_size)
speed = int(progress_size / (1024 * duration))
percent = min(int(count*block_size*100/total_size),100)
sys.stdout.write("\r...%d%%, %d MB, %d KB/s, %d seconds passed" %
(percent, progress_size / (1024 * 1024), speed, duration))
sys.stdout.flush()
urllib.request.urlretrieve(origin, fname, reporthook)
print('Download finished. Start extract!', origin)
if fname.endswith("tar.gz"):
tar = tarfile.open(fname, "r:gz")
tar.extractall()
tar.close()
elif fname.endswith("tar"):
tar = tarfile.open(fname, "r:")
tar.extractall()
tar.close()
def unpickle(file):
with open(file, 'rb') as fo:
dict = pickle.load(fo, encoding='bytes')
return dict
def load_data_one(file):
batch = unpickle(file)
data = batch[b'data']
labels = batch[b'labels']
print("Loading %s : %d." % (file, len(data)))
return data, labels
def load_data(files, data_dir, label_count):
global image_size, img_channels
data, labels = load_data_one(data_dir + '/' + files[0])
for f in files[1:]:
data_n, labels_n = load_data_one(data_dir + '/' + f)
data = np.append(data, data_n, axis=0)
labels = np.append(labels, labels_n, axis=0)
labels = np.array([[float(i == label) for i in range(label_count)] for label in labels])
data = data.reshape([-1, img_channels, image_size, image_size])
data = data.transpose([0, 2, 3, 1])
return data, labels
def prepare_data():
print("======Loading data======")
download_data()
#data_dir = './CAFIR-10_data/cifar-10-python/cifar-10-batches-py'
data_dir = './cifar-10-batches-py' # 数据集解压后的文件夹
image_dim = image_size * image_size * img_channels
meta = unpickle(data_dir + '/batches.meta')
label_names = meta[b'label_names']
label_count = len(label_names)
train_files = ['data_batch_%d' % d for d in range(1, 6)]
train_data, train_labels = load_data(train_files, data_dir, label_count)
test_data, test_labels = load_data(['test_batch'], data_dir, label_count)
print("Train data:", np.shape(train_data), np.shape(train_labels))
print("Test data :", np.shape(test_data), np.shape(test_labels))
print("======Load finished======")
print("======Shuffling data======")
indices = np.random.permutation(len(train_data))
train_data = train_data[indices]
train_labels = train_labels[indices]
print("======Prepare Finished======")
return train_data, train_labels, test_data, test_labels
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape, dtype=tf.float32)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool(input, k_size=1, stride=1, name=None):
return tf.nn.max_pool(input, ksize=[1, k_size, k_size, 1], strides=[1, stride, stride, 1],
padding='SAME', name=name)
def batch_norm(input):
return tf.contrib.layers.batch_norm(input, decay=0.9, center=True, scale=True, epsilon=1e-3,
is_training=train_flag, updates_collections=None)
# random crop 随机裁剪,下面的数据增强
def _random_crop(batch, crop_shape, padding=None):
oshape = np.shape(batch[0])
if padding:
oshape = (oshape[0] + 2*padding, oshape[1] + 2*padding)
new_batch = []
npad = ((padding, padding), (padding, padding), (0, 0))
for i in range(len(batch)):
new_batch.append(batch[i])
if padding:
new_batch[i] = np.lib.pad(batch[i], pad_width=npad,
mode='constant', constant_values=0)
nh = random.randint(0, oshape[0] - crop_shape[0])
nw = random.randint(0, oshape[1] - crop_shape[1])
new_batch[i] = new_batch[i][nh:nh + crop_shape[0],
nw:nw + crop_shape[1]]
return new_batch
def _random_flip_leftright(batch):
for i in range(len(batch)):
if bool(random.getrandbits(1)):
batch[i] = np.fliplr(batch[i])
return batch
def data_augmentation(batch):
batch = _random_flip_leftright(batch)
batch = _random_crop(batch, [32, 32], 4)
return batch
def data_preprocessing(x_train,x_test):
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train[:, :, :, 0] = (x_train[:, :, :, 0] - np.mean(x_train[:, :, :, 0])) / np.std(x_train[:, :, :, 0])
x_train[:, :, :, 1] = (x_train[:, :, :, 1] - np.mean(x_train[:, :, :, 1])) / np.std(x_train[:, :, :, 1])
x_train[:, :, :, 2] = (x_train[:, :, :, 2] - np.mean(x_train[:, :, :, 2])) / np.std(x_train[:, :, :, 2])
x_test[:, :, :, 0] = (x_test[:, :, :, 0] - np.mean(x_test[:, :, :, 0])) / np.std(x_test[:, :, :, 0])
x_test[:, :, :, 1] = (x_test[:, :, :, 1] - np.mean(x_test[:, :, :, 1])) / np.std(x_test[:, :, :, 1])
x_test[:, :, :, 2] = (x_test[:, :, :, 2] - np.mean(x_test[:, :, :, 2])) / np.std(x_test[:, :, :, 2])
return x_train, x_test
def learning_rate_schedule(epoch_num):
if epoch_num < 81:
return 0.1 # 0.01
elif epoch_num < 121:
return 0.01 #0.001
else:
return 0.001 #0.0001
def run_testing(sess, ep):
acc = 0.0
loss = 0.0
pre_index = 0
add = 1000
for it in range(10):
batch_x = test_x[pre_index:pre_index+add]
batch_y = test_y[pre_index:pre_index+add]
pre_index = pre_index + add
loss_, acc_ = sess.run([cross_entropy, accuracy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: False})
loss += loss_ / 10.0
acc += acc_ / 10.0
summary = tf.Summary(value=[tf.Summary.Value(tag="test_loss", simple_value=loss),
tf.Summary.Value(tag="test_accuracy", simple_value=acc)])
return acc, loss, summary
if __name__ == '__main__':
train_x, train_y, test_x, test_y = prepare_data()
train_x, test_x = data_preprocessing(train_x, test_x)
# define placeholder x, y_ , keep_prob, learning_rate
x = tf.placeholder(tf.float32,[None, image_size, image_size, 3])
y_ = tf.placeholder(tf.float32, [None, class_num])
keep_prob = tf.placeholder(tf.float32)
learning_rate = tf.placeholder(tf.float32)
train_flag = tf.placeholder(tf.bool)
# build_network
# 32x32x3
W_conv1_1 = tf.get_variable('conv1_1', shape=[3, 3, 3, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_1 = bias_variable([64])
output = tf.nn.relu(batch_norm(conv2d(x, W_conv1_1) + b_conv1_1))
W_conv1_2 = tf.get_variable('conv1_2', shape=[3, 3, 64, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_2 = bias_variable([64])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv1_2) + b_conv1_2))
output = max_pool(output, 2, 2, "pool1") # 16x16x64
W_conv2_1 = tf.get_variable('conv2_1', shape=[3, 3, 64, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_1 = bias_variable([128])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv2_1) + b_conv2_1))
W_conv2_2 = tf.get_variable('conv2_2', shape=[3, 3, 128, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_2 = bias_variable([128])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv2_2) + b_conv2_2))
output = max_pool(output, 2, 2, "pool2") # 8x8x128
W_conv3_1 = tf.get_variable('conv3_1', shape=[3, 3, 128, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_1 = bias_variable([256])
output = tf.nn.relu( batch_norm(conv2d(output,W_conv3_1) + b_conv3_1))
W_conv3_2 = tf.get_variable('conv3_2', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_2 = bias_variable([256])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv3_2) + b_conv3_2))
W_conv3_3 = tf.get_variable('conv3_3', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_3 = bias_variable([256])
output = tf.nn.relu( batch_norm(conv2d(output, W_conv3_3) + b_conv3_3))
output = max_pool(output, 2, 2, "pool3") # 4x4x256
W_conv4_1 = tf.get_variable('conv4_1', shape=[3, 3, 256, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_1 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_1) + b_conv4_1))
W_conv4_2 = tf.get_variable('conv4_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_2 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_2) + b_conv4_2))
W_conv4_3 = tf.get_variable('conv4_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_3 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv4_3) + b_conv4_3))
output = max_pool(output, 2, 2) # 2x2x512
W_conv5_1 = tf.get_variable('conv5_1', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_1 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_1) + b_conv5_1))
W_conv5_2 = tf.get_variable('conv5_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_2 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_2) + b_conv5_2))
W_conv5_3 = tf.get_variable('conv5_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_3 = bias_variable([512])
output = tf.nn.relu(batch_norm(conv2d(output, W_conv5_3) + b_conv5_3))
##output = max_pool(output, 2, 2) # 这里应该要池化 ######3
# output = tf.contrib.layers.flatten(output)
output = tf.reshape(output, [-1, 2*2*512])
##output = tf.reshape(output, [-1, 1 * 1 * 512])
##W_fc1 = tf.get_variable('fc1', shape=[1 * 1 * 512, 4096], initializer=tf.contrib.keras.initializers.he_normal())
W_fc1 = tf.get_variable('fc1', shape=[2 * 2 * 512, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc1 = bias_variable([4096])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc1) + b_fc1) )
output = tf.nn.dropout(output, keep_prob)
W_fc2 = tf.get_variable('fc7', shape=[4096, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc2 = bias_variable([4096])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc2) + b_fc2))
output = tf.nn.dropout(output, keep_prob)
W_fc3 = tf.get_variable('fc3', shape=[4096, 10], initializer=tf.contrib.keras.initializers.he_normal())
b_fc3 = bias_variable([10])
output = tf.nn.relu(batch_norm(tf.matmul(output, W_fc3) + b_fc3))
# output = tf.reshape(output,[-1,10])
# loss function: cross_entropy
# train_step: training operation
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=output))
l2 = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
train_step = tf.train.MomentumOptimizer(learning_rate, momentum_rate, use_nesterov=True).\
minimize(cross_entropy + l2 * weight_decay)
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# initial an saver to save model
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(log_save_path,sess.graph)
# epoch = 164
# make sure [bath_size * iteration = data_set_number]
for ep in range(1, total_epoch+1):
lr = learning_rate_schedule(ep)
pre_index = 0
train_acc = 0.0
train_loss = 0.0
start_time = time.time()
print("\n epoch %d/%d:" % (ep, total_epoch))
for it in range(1, iterations+1):
batch_x = train_x[pre_index:pre_index+batch_size]
batch_y = train_y[pre_index:pre_index+batch_size]
batch_x = data_augmentation(batch_x)
_, batch_loss = sess.run([train_step, cross_entropy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: dropout_rate,
learning_rate: lr, train_flag: True})
batch_acc = accuracy.eval(feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: True})
train_loss += batch_loss
train_acc += batch_acc
pre_index += batch_size
if it == iterations:
train_loss /= iterations
train_acc /= iterations
loss_, acc_ = sess.run([cross_entropy, accuracy],
feed_dict={x: batch_x, y_: batch_y, keep_prob: 1.0, train_flag: True})
train_summary = tf.Summary(value=[tf.Summary.Value(tag="train_loss", simple_value=train_loss),
tf.Summary.Value(tag="train_accuracy", simple_value=train_acc)])
val_acc, val_loss, test_summary = run_testing(sess, ep)
summary_writer.add_summary(train_summary, ep)
summary_writer.add_summary(test_summary, ep)
summary_writer.flush()
print("iteration: %d/%d, cost_time: %ds, train_loss: %.4f, "
"train_acc: %.4f, test_loss: %.4f, test_acc: %.4f"
% (it, iterations, int(time.time()-start_time), train_loss, train_acc, val_loss, val_acc))
else:
print("iteration: %d/%d, train_loss: %.4f, train_acc: %.4f"
% (it, iterations, train_loss / it, train_acc / it))
save_path = saver.save(sess, model_save_path)
print("Model saved in file: %s" % save_path)
目前，在将代码改为tensorflow2.0版本，后面有时间在写一篇关于2.0的训练代码。
参考博文网址：https://blog.csdn/GOGO_YAO/article/details/80348200;