原文连接 https://blog.csdn.net/yutingzhaomeng/article/details/81708261
本文总结tensorflow使用的相关方法，包括：
0、定义网络输入
1、如何利用tensorflow在已有网络入resnet基础上搭建自己的网络结构
2、如何添加自己的网络层
3、如何导入已有模块入resnet全连接层之前部分的参数
4、定义网络损失
5、定义优化算子以及衰减优化算子
6、预测网络输出
7、保存网络模型
8、自定义生成训练batch
9、训练网络
10、利用tensorboard可视化训练过程
0、定义网络输入
inputs = tf.placeholder(tf.float32, [None, 224, 224, 3], name='inputs')
labels = tf.placeholder(tf.int32, [None], name='lables')
is_training = tf.placeholder(tf.bool, name='is_training')
这里inputs表示输入数据，labels表示对应的label，is_training主要用于区分如drop和batchnorm层的训练测试阶段。
1、如何利用tensorflow在已有网络入resnet基础上搭建自己的网络结构
with slim.arg_scope(nets.resnet_v1.resnet_arg_scope()):
if config.TRAIN.net_layer == '50':
logits, endpoints = nets.resnet_v1.resnet_v1_50(inputs, num_classes=None, is_training=is_training)
if config.TRAIN.net_layer == '101':
logits, endpoints = nets.resnet_v1.resnet_v1_101(inputs, num_classes=None, is_training=is_training)
if config.TRAIN.net_layer == '152':
logits, endpoints = nets.resnet_v1.resnet_v1_152(inputs, num_classes=None, is_training=is_training)
以resnet为例，logits表示bottleneck特征，num_classes设置为None表示取bottleneck特征。
2、如何添加自己的网络层
with tf.variable_scope('Logits'):
logits = tf.squeeze(logits, axis=[1,2])
logits = slim.dropout(logits, keep_prob=0.5, scope='scope')
logits = slim.fully_connected(logits, num_outputs=config.DATASET.num_classes, activation_fn=None, scope='fc')
这里有一个scope，后面我们会发现，主要用来区别resnet已有参数，squeeze用于将1*1*512的特征拉伸为向量，我们添加dropout层和全连接层。
3、如何导入已有模块入resnet全连接层之前部分的参数
checkpoint_exclude_scopes = 'Logits'
exclusions = None
if checkpoint_exclude_scopes:
exclusions = [scope.strip() for scope in checkpoint_exclude_scopes.split(',')]
variables_to_restore = []
for var in slim.get_model_variables():
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
if not excluded:
variables_to_restore.append(var)
logits scope下的变量我们不考虑，其他参数restore恢复。
4、定义网络损失
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=logits))
5、定义优化算子以及衰减优化算子
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.0001)
train_step = optimizer.minimize(loss)
batch = config.TRAIN.batch_size
sample_size = len(os.listdir(config.DATASET.image_root))
global_step = tf.Variable(0)
learning_rate = tf.train.exponential_decay(1e-4, global_step,
decay_steps=4 * sample_size / batch, decay_rate=0.98,
staircase=True)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
上面的表示正常定义优化算子，下面的表示衰减优化算子。其中，batch表示每个batch样本数，sample_size即样本数，global_step用于获取当前iteration，sample_size / batch即每个epoch包含的iteration数目，计算衰减时，每一个decay_steps降低一次学习率。learning_rate_current = learning_rate_start * dacay_rate ** (global_step / decay_steps)。
6、预测网络输出
logits = tf.nn.softmax(logits, name='logits')
classes = tf.argmax(logits, axis=1, name='classes')
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.cast(classes, tf.int32), labels), tf.float32))
7、保存网络模型
init = tf.global_variables_initializer()
saver_restore = tf.train.Saver(var_list=variables_to_restore)
saver = tf.train.Saver(tf.global_variables())
8、自定义生成训练batch
images, truths, valid_imgs, valid_trus = get_batch()
def get_label(xml_path):
tree = ET.parse(xml_path)
objs = tree.findall('object')
objs = [obj for obj in objs if 'b' in obj.find('name').text] # select all pointer pannels
if not len(objs) == 1:
return [[], []]
obj = objs[0] # suppose there is only one pannel, otherwise use center selection
label = str(float(obj.find('name').text.split('b')[-1]))
return [label]
def get_list():
image_list = []
label_list = []
for file in os.listdir(config.DATASET.image_root):
image_label = get_label(os.path.join(config.DATASET.label_root,file.split('.jpg')[0]+'.xml'))
if len(image_label) > 1:
continue
else:
image_label = image_label[0]
if image_label in config.DATASET.range_dict.keys():
label_list.append(config.DATASET.range_dict[image_label])
else:
label_list.append(len(config.DATASET.range_dict))
image_list.append(os.path.join(config.DATASET.image_root,file))
valid_num = int(len(image_list)*config.DATASET.valid_ratio)
train_list = image_list[valid_num:]
valid_list = image_list[:valid_num]
train_label = label_list[valid_num:]
valid_label = label_list[:valid_num]
return train_list, train_label, valid_list, valid_label
def process_batch(input_quene):
label = input_quene[1]
image = tf.read_file(input_quene[0])
image = tf.image.decode_jpeg(image, channels=3)
image = tf.image.resize_image_with_crop_or_pad(image, config.DATASET.width, config.DATASET.height)
image = tf.image.per_image_standardization(image)
image_batch, label_batch = tf.train.batch([image, label], batch_size=config.TRAIN.batch_size,
capacity=config.TRAIN.capacity, num_threads=config.TRAIN.num_threads)
label_batch = tf.reshape(label_batch, [config.TRAIN.batch_size])
image_batch = tf.cast(image_batch, tf.float32)
return image_batch, label_batch
def get_batch():
train_image_list, train_label_list, valid_image_list, valid_label_list = get_list()
input_quene = tf.train.slice_input_producer([train_image_list, train_label_list])
trian_image_batch, trian_label_batch = process_batch(input_quene)
valid_quene = tf.train.slice_input_producer([valid_image_list, valid_label_list])
valid_image_batch, valid_label_batch = process_batch(valid_quene)
return trian_image_batch, trian_label_batch, valid_image_batch, valid_label_batch
9、训练网络
with tf.Session(config=tfConfig) as sess:
sess.run(init)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# =============================Import Pretrained Parameter=========================== #
saver_restore.restore(sess, config.TRAIN.model_path)
# ================================TensorBoard Related================================ #
tf.summary.image('inputs',inputs)
tf.summary.scalar('loss',loss)
tf.summary.scalar('accuracy',accuracy)
tf.summary.scalar('learning rate', learning_rate)
merged_summary_op = tf.summary.merge_all()
if os.path.exists(os.path.join(config.TRAIN.log_path, 'train')):
shutil.rmtree(os.path.join(config.TRAIN.log_path, 'train'))
if os.path.exists(os.path.join(config.TRAIN.log_path, 'valid')):
shutil.rmtree(os.path.join(config.TRAIN.log_path, 'valid'))
train_writer = tf.summary.FileWriter(os.path.join(config.TRAIN.log_path, 'train'), sess.graph)
valid_writer = tf.summary.FileWriter(os.path.join(config.TRAIN.log_path, 'valid'))
for i in range(config.TRAIN.num_iterations):
images_, truths_ = sess.run([images, truths])
valid_imgs_, valid_trus_ = sess.run([valid_imgs, valid_trus])
summary_str, _, loss_, acc_ = sess.run([merged_summary_op, train_step, loss, accuracy],
feed_dict={inputs: images_, labels: truths_, is_training: True})
valid_str, vloss, vacc = sess.run([merged_summary_op, loss, accuracy],
feed_dict={inputs: valid_imgs_, labels: valid_trus_, is_training: False})
print('Step: {}, Loss: {:.4f}, Accuracy: {:.4f}, Valid Loss: {:.4f}, Valid Accuracy: {:.4f}'.format(i+1, loss_, acc_, vloss, vacc))
# if (i+1) % 1000 == 0:
# saver.save(sess, config.TRAIN.save_path)
# print('save mode to {}'.format(config.TRAIN.save_path))
# summary_str = sess.run(merged_summary_op)
train_writer.add_summary(summary_str, i)
valid_writer.add_summary(valid_str, i)
coord.request_stop()
coord.join(threads)
10、利用tensorboard可视化训练过程
tf.summary.image('inputs',inputs)
tf.summary.scalar('loss',loss)
tf.summary.scalar('accuracy',accuracy)
tf.summary.scalar('learning rate', learning_rate)
merged_summary_op = tf.summary.merge_all()
if os.path.exists(os.path.join(config.TRAIN.log_path, 'train')):
shutil.rmtree(os.path.join(config.TRAIN.log_path, 'train'))
if os.path.exists(os.path.join(config.TRAIN.log_path, 'valid')):
shutil.rmtree(os.path.join(config.TRAIN.log_path, 'valid'))
train_writer = tf.summary.FileWriter(os.path.join(config.TRAIN.log_path, 'train'), sess.graph)
valid_writer = tf.summary.FileWriter(os.path.join(config.TRAIN.log_path, 'valid'))