def predict_and_save_pics_model():
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
prediction, end_points = model(images=x,
is_train=is_train,
size=input_shape,
l2_reg=0.0001)
# 打印模型结构
dk.print_model_struct(end_points)
# prediction = tf.reshape(prediction, labels_shape)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 恢复model
saver = dk.restore_model(sess, ckpt, restore_model=restore_model)
# 显示参数量
dk.show_parament_numbers()
# 测试loop
start_epoch = 0
index = 0
for epoch_n in range(start_epoch, test_opoch):
for n_batch in range(n_batch_test):
batch_x, batch_y = sess.run([train_x,
train_y]) # 取出一个batchsize的图片。
batch_x = batch_x / 255.0
# 3、预测输出
since = time.time()
predict = sess.run(prediction, feed_dict={x: batch_x})
seconds = time.time() - since
# 4、预测图转二值化图（非0即1）
predict[predict >= 0.5] = 1
predict[predict < 0.5] = 0
# 5、把batch图片分割一张张图片
batch_pre_list = np.split(predict, batch_size, axis=0)
pre_pics_list = np.squeeze(batch_pre_list, axis=0)
for img, label, predict in zip(batch_x, batch_y,
pre_pics_list):
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
label = cv2.cvtColor(label, cv2.COLOR_GRAY2RGB)
predict = cv2.cvtColor(predict, cv2.COLOR_GRAY2RGB)
hstack = np.hstack((img, label, predict))
hstack = cv2.resize(hstack, (512, 512))
save_name = '{}.jpg'.format(index)
save_name = os.path.join(predict_pics_save, save_name)
cv2.imshow('hstack', hstack)
cv2.waitKey(500)
cv2.imwrite(save_name, hstack * 255)
index += 1
print(save_name)
dk.stop_threads(coord, threads)
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)
def evaluate_result():
with tf.Session(config=session_config) as sess:
test_x, test_y = create_inputs(is_train)
# test_y = tf.one_hot(test_y, depth=class_num, axis=1, dtype=tf.int32)
# 构建网络
x = tf.placeholder(tf.float32,
shape=[
input_shape[0], input_shape[1], input_shape[2],
input_shape[3]
])
y = tf.placeholder(tf.float32, shape=[input_shape[0], class_num])
prediction, endpoint = model(x, input_shape, class_num)
# 打印模型结构
dk.print_model_struct(endpoint)
# 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 = dk.restore_model(sess, ckpt, restore_model=restore_model)
# 显示参数量
dk.show_parament_numbers()
start_epoch = 0
acc_list = []
if is_train:
print('train_number:', train_number)
n_batch_train = int(train_number // batch_size)
n_batch_total = n_batch_train
else:
print('test_number:', test_number)
n_batch_test = int(test_number // batch_size)
n_batch_total = n_batch_test
for epoch_n in range(start_epoch, epoch):
for n_batch in range(n_batch_total):
batch_x, batch_y = sess.run([test_x, test_y])
# 训练一个step
acc_value = sess.run(accuracy,
feed_dict={
x: batch_x,
y: batch_y
})
# 显示结果batch_size
print('epoch_n:', epoch_n, ' n_batch:', n_batch, ' acc_value:',
acc_value)
acc_list.append(acc_value)
result = np.mean(acc_list)
print('final result: ', result)
dk.stop_threads(coord, threads)
os.makedirs(save_plot_curve_dir,exist_ok=True)
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
prediction = model(images=x, is_train=True, size=input_shape, l2_reg=0.0001)
prediction = tf.reshape(prediction, labels_shape)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 恢复model
saver = dk.restore_model(sess, ckpt, restore_model=restore_model)
# 显示参数量
dk.show_parament_numbers()
# 列表初始化
AUC_ROC_list,AUC_prec_rec_list,accuracy_list,specificity_list,sensitivity_list,\
precision_list,jaccard_index_list,F1_score_list,all_list,mean_list\
=[],[],[],[],[],[],[],[],[],[]
# 测试loop
start_epoch= 0
index = 0
for epoch_n in range(start_epoch, test_opoch):
for n_batch in range(n_batch_test):
batch_x,batch_y = sess.run([train_x,train_y]) # 取出一个batchsize的图片。
batch_x = batch_x / 255.0
# 3、预测输出
since = time.time()
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)
# 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)
# 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(start_epoch,epoch):
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)
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)
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 = cnn_mnist(x, keep_prob=0.8)
# 求loss
loss = dk.cross_entropy_loss(prediction, y)
# 设置优化器
global_step, train_step = dk.set_optimizer(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 = 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_n in range(epoch):
pre_index = 0
for batch 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
})
def evaluate_by_sklearn_model():
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
prediction, end_points = model(images=x,
is_train=True,
size=input_shape,
l2_reg=0.0001)
prediction = tf.reshape(prediction, labels_shape)
# 打印模型结构
dk.print_model_struct(end_points)
# 初始化变量
coord, threads = dk.init_variables_and_start_thread(sess)
# 恢复model
saver = dk.restore_model(sess, ckpt, restore_model=restore_model)
# 显示参数量
dk.show_parament_numbers()
# 列表初始化
AUC_ROC_list,AUC_prec_rec_list,accuracy_list,specificity_list,sensitivity_list,\
precision_list,jaccard_index_list,F1_score_list,all_list,mean_list\
=[],[],[],[],[],[],[],[],[],[]
# 测试loop
start_epoch = 0
index = 0
for epoch_n in range(start_epoch, test_opoch):
for n_batch in range(n_batch_test):
batch_x, batch_y = sess.run([train_x,
train_y]) # 取出一个batchsize的图片。
batch_x = batch_x / 255.0
# 3、预测输出
since = time.time()
predict = sess.run(prediction, feed_dict={x: batch_x})
seconds = time.time() - since
# 4、预测图转二值化图（非0即1）,加了这步性能指标会下降2%左右
predict[predict >= 0.5] = 1
predict[predict < 0.5] = 0
# 5、把batch图片分割一张张图片
batch_pre_list = np.split(predict, batch_size, axis=0)
pre_pics_list = np.squeeze(batch_pre_list, axis=0)
for label, pre in zip(batch_y, pre_pics_list):
index += 1
y_scores = pre.reshape(-1, 1)
y_true = label.reshape(-1, 1)
print(
'######################### start ####################################'
)
# 1、画ROC曲线
AUC_ROC = esk.plot_roc_curve(y_true,
y_scores,
save_plot_curve_dir,
curve_name=str(index))
AUC_ROC_list.append(AUC_ROC)
#2、画P_R-curve曲线
AUC_prec_rec = esk.plot_precision_recall_curve(
y_true,
y_scores,
save_plot_curve_dir,
curve_name=str(index))
#3、Confusion matrix
y_pred_binary = esk.convert_to_binary(
shape=y_scores.shape[0], y_scores=y_scores)
accuracy, specificity, sensitivity, precision \
= esk.plot_confusion_matrix(y_true, y_pred_binary)
#4、Jaccard similarity index
jaccard_index = esk.get_jaccard_index(
y_true, y_pred_binary)
#5、F1 score
F1_score = esk.get_F1_score(y_true, y_pred_binary)
print(
'######################### end ####################################'
)
# 保存结果
AUC_prec_rec_list.append(AUC_prec_rec)
accuracy_list.append(accuracy)
specificity_list.append(specificity)
sensitivity_list.append(sensitivity)
precision_list.append(precision)
jaccard_index_list.append(jaccard_index)
F1_score_list.append(F1_score)
#1、评估数据存进列表中
all_list = [AUC_ROC_list, AUC_prec_rec_list, accuracy_list, specificity_list \
, sensitivity_list, precision_list, jaccard_index_list, F1_score_list]
name_list = [
'AUC_ROC', 'AUC_prec_rec', 'accuracy', 'specificity',
'sensitivity', 'precision', 'jaccard_index', 'F1_score'
]
# 2、panda保存所有图片的评估值到CSV文件
esk.save_all_pics_value(name_list, all_list, save_list_csv)
#3、panda保存平均值到CSV文件
esk.save_mean_value(name_list, all_list, save_mean_csv)
#4、结束
dk.stop_threads(coord, threads)
def predict_and_save_tensor_feature_map_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)
# 构建网络
prediction, end_points = model(images=x,
is_train=is_train,
size=input_shape,
l2_reg=0.0001)
# 打印模型结构
dk.print_model_struct(end_points)
# 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=restore_model)
# 显示参数量
dk.show_parament_numbers()
# 测试loop
n_batch_index = 0
n_pic_index = 0
channel_index = 0
for n_batch in range(n_batch_test):
n_batch_index += 1
batch_x, batch_y = sess.run([train_x,
train_y]) # 取出一个batchsize的图片。
batch_x = batch_x / 255.0
# 3、预测输出一个张量
for key, value in end_points.items():
#保存的路径
save_root = os.path.join(predict_tensor_feature_map, key)
os.makedirs(save_root, exist_ok=True)
predict_end_points = sess.run(value, feed_dict={x: batch_x})
shape = predict_end_points.shape
if len(shape) == 4:
batch_size, width, height, channel = shape
elif len(shape) == 5:
batch_size, width, height, channel_1, channel_2 = shape
shape = (batch_size, width, height, channel_1 * channel_2)
predict_end_points = np.reshape(predict_end_points, shape)
channel = channel_1 * channel_2
else:
raise Exception('f**k!')
#切分batch_size
batch_size_list = np.split(predict_end_points,
batch_size,
axis=0)
pics_list = [] # 清空pics_list
for each_b in batch_size_list: #循环一个batch_size里面的每张图片图片
pic_channel_list = []
#一张图片切分成channel个，把每个featu map装进pics_list里面
channel_list = np.split(each_b, channel, axis=3)
for each_c in channel_list: # 切分channel（显示每个channel的图片）
each_c = np.reshape(each_c,
(width, height, 1)) #得到一张图片
# 显示方式1：二值化
# 4、预测图转二值化图（非0即1） ，经过试验tensor阈值为0最佳
# each_c[each_c >= 0] = 255
# each_c[each_c < 0] = 0
# pics_list.extend([each_c])
# 显示方式2：归一化并转成灰度图处理
each_c = MaxMinNormalization(each_c,
Max=np.max(each_c),
Min=np.min(each_c))
# each_c = cv2.cvtColor(each_c, cv2.COLOR_GRAY2RGB)
# pics_list.extend([each_c])
# 显示方式3：
# each_c = Z_ScoreNormalization(each_c, mu=np.average(each_c), sigma=np.std(each_c))
# each_c = cv2.cvtColor(each_c, cv2.COLOR_GRAY2RGB)
pic_channel_list.extend([each_c]) #一张图片所有的channel
pics_list.append(pic_channel_list) #把pic_channel_list添加进去
for pic_channel_list in pics_list: #第几张图片
n_pic_index += 1
new_dir = '{}_{}'.format(n_batch_index, n_pic_index)
save_dir = os.path.join(save_root, new_dir)
os.makedirs(save_dir, exist_ok=True)
for pic in pic_channel_list: #第几个通道
channel_index += 1
# 保存的文件名
save_name = '{}.jpg'.format(channel_index)
# 保存的整个路径
save_name = os.path.join(save_dir, save_name)
cv2_show = True
if cv2_show == True:
cv2.imshow('pic', pic)
cv2.imwrite(save_name, pic * 255)
cv2.waitKey(500)
else:
plt_imshow_1_pics(pic, save_name)
time.sleep(1)
print(save_name)
channel_index = 0
n_pic_index = 0
n_batch_index = 0
dk.stop_threads(coord, threads)
# coord.request_stop()
# coord.join(threads)
########################废弃代码
# 4、预测图转二值化图（非0即1）
# predict[predict>=0.5] =1
# predict[predict< 0.5] =0
# 5、把batch图片分割一张张图片
# batch_pre_list = np.split(predict, batch_size, axis=0)
# pre_pics_list = np.squeeze(batch_pre_list, axis=0)
# for img, label, predict in zip(batch_x, batch_y, pre_pics_list):
# img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# label = cv2.cvtColor(label, cv2.COLOR_GRAY2RGB)
# predict = cv2.cvtColor(predict, cv2.COLOR_GRAY2RGB)
# hstack = np.hstack((img, label,predict))
# hstack = cv2.resize(hstack,(512,512))
#
# save_name = '{}.jpg'.format(index)
# save_name = os.path.join(predict_pics_save, save_name)
# cv2.imshow('hstack',hstack)
# cv2.waitKey(500)
# cv2.imwrite(save_name, hstack*255)
# index += 1
# print(save_name)
