def train_mnist_model():
with tf.Session(config=session_config) as sess:
mnist = create_inputs.read_data_sets(dataset_path, one_hot=True)
# 计算有多少批次
n_batch = mnist.train.num_examples // batch_size
# 构建网络
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, class_num])
prediction, _ = model(x, input_shape, class_num)
# 求loss
loss = dk.cross_entropy_loss(prediction, y)
# 设置优化器
global_step, train_step = dk.set_optimizer(
lr_range, num_batches_per_epoch=n_batch, loss=loss)
# 求acc
accuracy = dk.get_acc(prediction, y)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 恢复model
saver, start_epoch = dk.restore_model(sess, ckpt, restore_model=False)
# 设置训练日志
summary_dict = {'loss': loss, 'accuracy': accuracy}
summary_writer, summary_op = dk.set_summary(sess, logdir, summary_dict)
# epoch=50
for epoch_th in range(epoch):
for bat in range(n_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
# 训练一个step
_, loss_value, acc_value, summary_str, step = sess.run(
[train_step, loss, accuracy, summary_op, global_step],
feed_dict={
x: batch_xs,
y: batch_ys
})
# 显示结果
dk.print_message(epoch_th, bat, n_batch, step, loss_value,
acc_value)
# 保存summary
if (step + 1) % 20 == 0:
summary_writer.add_summary(summary_str, step)
# 保存model
if (((epoch_th + 1) % 50)) == 0:
print('saving movdel.......')
saver.save(sess,
os.path.join(ckpt,
'model_{}.ckpt'.format(epoch_th)),
global_step=global_step)
dk.stop_threads(coord, threads)
if __name__== '__main__':
with tf.Session(config = session_config) as sess:
# 入口
train_x, train_y = create_inputs(is_train)
# train_y = tf.reshape(train_y,labels_shape)
x = tf.placeholder(tf.float32, shape=input_shape)
y = tf.placeholder(tf.float32, shape=labels_shape)
# 构建网络和预测
prediction = model(images= x, is_train =is_train,size= input_shape,l2_reg =0.0001 )
# 求loss
# loss = dk.cross_entropy_loss(prediction, y)
the_loss = get_loss('bce_dice')
loss = the_loss(y, prediction,labels_shape_vec)
# 设置优化器
global_step, train_step = dk.set_optimizer(num_batches_per_epoch=n_batch_train, loss=loss)
# 求dice_hard，不合适用acc
dice_hard = dk.dice_hard(y, prediction, threshold=0.5, axis=[1, 2, 3], smooth=1e-5)
# accuracy = dk.get_acc(prediction, y)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 设置训练日志
summary_dict = {'loss':loss,'dice_hard':dice_hard}
summary_writer, summary_op = dk.set_summary(sess,logdir,summary_dict)
# 恢复model
saver,start_epoch = dk.restore_model(sess, ckpt, restore_model=restore_model) # 显示参数量
dk.show_parament_numbers()
# 若恢复model，则重新计算start_epoch继续
# start_epoch = 0
# if restore_model:
# step = sess.run(global_step)
def train_cifar10_model():
with tf.Session(config=session_config) as sess:
#入口
train_x, train_y = create_inputs(is_train)
x = tf.placeholder(tf.float32, shape=input_shape)
y = tf.placeholder(tf.float32, shape=label_shape)
# 构建网络
prediction, _ = model(x, input_shape, class_num)
# 求loss
loss = dk.cross_entropy_loss(prediction, y)
# 设置优化器
global_step, train_step = dk.set_optimizer(
lr_range, num_batches_per_epoch=n_batch_train, loss=loss)
# 求acc
accuracy = dk.get_acc(prediction, y)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 设置训练日志
summary_dict = {'loss': loss, 'accuracy': accuracy}
summary_writer, summary_op = dk.set_summary(sess, logdir, summary_dict)
# 恢复model
saver, start_epoch = dk.restore_model(sess, ckpt, restore_model=False)
# 显示参数量
dk.show_parament_numbers()
start_epoch = 0
if restore_model:
step = sess.run(global_step)
start_epoch = int(
step / n_batch_train / save_epoch_n) * save_epoch_n
# 训练loop
total_step = n_batch_train * epoch
for epoch_n in range(epoch):
pre_index = 0
since = time.time()
acc_value_list = []
for n_batch in range(n_batch_train):
if pre_index + batch_size < 50000:
batch_x = train_x[pre_index:pre_index + batch_size]
batch_y = train_y[pre_index:pre_index + batch_size]
else:
batch_x = train_x[pre_index:]
batch_y = train_y[pre_index:]
# 训练一个step
_, loss_value, acc_value, summary_str, step = sess.run(
[train_step, loss, accuracy, summary_op, global_step],
feed_dict={
x: batch_x,
y: batch_y
})
# 显示结果
dk.print_message(epoch_n, n_batch, n_batch_train, step,
loss_value, acc_value)
# 保存summary
if (step + 1) % 20 == 0:
summary_writer.add_summary(summary_str, step)
pre_index += batch_size
# 保存结果
acc_value_list.append(acc_value)
# 显示进度、耗时、最小最大平均值
seconds_mean = (time.time() - since) / n_batch_train
dk.print_progress_and_time_massge(seconds_mean, step, total_step,
acc_value_list)
# 保存model
if (((epoch_n + 1) % save_epoch_n)) == 0:
print('epoch_n :{} saving movdel.......'.format(epoch_n))
saver.save(sess,
os.path.join(ckpt, 'model_{}.ckpt'.format(epoch_n)),
global_step=global_step)
dk.stop_threads(coord, threads)
def train_model():
# 代码初始化
n_batch_train = int(train_data_number // batch_size)
print('n_batch_train: ', n_batch_train)
os.makedirs(ckpt, exist_ok=True)
session_config = dk.set_gpu()
with tf.Session(config=session_config) as sess:
#如果使用tensorlfow1的debug神器（主要用于查出哪里有inf或nan，不能在pycharm运行调试程序，只能在xshell里面运行）
if use_tensoflow_debug:
sess = tfdbg.LocalCLIDebugWrapperSession(sess)
sess.add_tensor_filter("has_inf_or_nan", tfdbg.has_inf_or_nan)
#然后在xshell里面运行run -f has_inf_or_nan
# 一旦inf / nan出现，界面现实所有包含此类病态数值的张量，按照时间排序。所以第一个就最有可能是最先出现inf / nan的节点。
# 可以用node_info, list_inputs等命令进一步查看节点的类型和输入，来发现问题的缘由。
#教程https://blog.csdn.net/tanmx219/article/details/82318133
# 入口
train_x, train_y = create_inputs(is_train)
x = tf.placeholder(tf.float32, shape=input_shape)
y = tf.placeholder(tf.float32, shape=labels_shape)
# 构建网络和预测
prediction, endpoint = model(images=x,
is_train=is_train,
size=input_shape,
l2_reg=0.0001)
# 打印模型结构
dk.print_model_struct(endpoint)
# 求loss
the_loss = get_loss(choose_loss)
loss = the_loss(y, prediction, labels_shape_vec)
# 设置优化器
global_step, train_step = dk.set_optimizer(
lr_range=lr_range, num_batches_per_epoch=n_batch_train, loss=loss)
# 求dice_hard，不合适用acc
dice_hard = dk.dice_hard(y,
prediction,
threshold=0.5,
axis=[1, 2, 3],
smooth=1e-5)
# dice_hard = dk.iou_metric(prediction, y)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 设置训练日志
summary_dict = {'loss': loss, 'dice_hard': dice_hard}
summary_writer, summary_op = dk.set_summary(sess, logdir, summary_dict)
# 恢复model
saver, start_epoch = dk.restore_model(sess,
ckpt,
restore_model=restore_model)
# 显示参数量
dk.show_parament_numbers()
# 训练loop
total_step = n_batch_train * epoch
for epoch_n in range(start_epoch, epoch):
dice_hard_value_list = [] #清空
since = time.time()
for n_batch in range(n_batch_train):
batch_x, batch_y = sess.run([train_x, train_y])
########################## 数据增强 ###################################
batch_x = batch_x / 255.0 # 归一化，加了这句话loss值小了几十倍
batch_x, batch_y = augmentImages(batch_x, batch_y)
########################## end #######################################
# 训练一个step
_, loss_value, dice_hard_value, summary_str, step = sess.run(
[train_step, loss, dice_hard, summary_op, global_step],
feed_dict={
x: batch_x,
y: batch_y
})
# 显示结果batch_size
dk.print_effect_message(epoch_n, n_batch, n_batch_train,
loss_value, dice_hard_value)
# 保存summary
if (step + 1) % 20 == 0:
summary_writer.add_summary(summary_str, step)
# 保存结果
dice_hard_value_list.append(dice_hard_value)
# 显示进度、耗时、最小最大平均值
seconds_mean = (time.time() - since) / n_batch_train
dk.print_progress_and_time_massge(seconds_mean, step, total_step,
dice_hard_value_list)
# 保存model
if (((epoch_n + 1) % save_epoch_n)) == 0:
print('epoch_n :{} saving movdel.......'.format(epoch_n))
saver.save(sess,
os.path.join(ckpt, 'model_{}.ckpt'.format(epoch_n)),
global_step=global_step)
dk.stop_threads(coord, threads)