def run_eval(val_loop=2,
number_of_classes=5,
batch_size=40,
sigma=0.6,
checkpoint_dir=None,
dataset_dir=None):
#定义输入变量
tf.reset_default_graph()
with tf.Graph().as_default() as g:
# 生成训练数据
img_val, label_val = gen_data_batch(dataset_dir=dataset_dir,
batch_size=batch_size,
Train=False)
# 推理、训练、准确度
x_img_in = tf.placeholder(tf.float32, [batch_size, 224, 224, 3])
x_label = tf.placeholder(tf.int64, [batch_size])
label_hot = tf.one_hot(x_label, number_of_classes)
label_advers = number_of_classes - 1 - x_label
label_advers_hot = tf.one_hot(label_advers, number_of_classes)
layer_name1 = 'resnet_v2_152/block4'
target_conv_layer, target_grad_ac, target_grad_yc, logits_cl, end_points = forward_tran(
x_img=x_img_in,
number_of_classes=number_of_classes,
layer_name=layer_name1,
Training=False)
conv_in = end_points[layer_name1]
mask_conv_imgs,_=batch_mask_images(batch_size=batch_size,img_conv = conv_in,target_conv_layer=target_conv_layer,\
target_grad_ac=target_grad_ac,target_grad_yc=target_grad_yc,sigma=sigma)
_, _, _, logits_adver, _ = forward_tran_advers(
x_img=mask_conv_imgs,
number_of_classes=number_of_classes,
Training=False)
accuracy = evaluation(logits_cl, label_hot)
accuracy_adver = evaluation(logits_adver, label_advers_hot)
prob_max = tf.argmax(tf.nn.softmax(logits_cl), 1)
prob_adver_max = tf.argmax(tf.nn.softmax(logits_adver), 1)
#
config = tf.ConfigProto() # 配置GPU参数
config.gpu_options.allow_growth = True # 动态分配GPU资源
#config.gpu_options.per_process_gpu_memory_fraction = 0.85 # 占用GPU90%的显存
sess = tf.Session(config=config)
#
start_time = time.time()
with sess: #开始一个会话
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('checkpoint restored from [{0}]'.format(
ckpt.model_checkpoint_path))
print('model_checkpoint_path Loading success')
else:
print('No checkpoint before training')
pass
accuracy_list = []
for i in range(val_loop):
_img_val, _label_val = sess.run([img_val, label_val])
feed_vars = {x_img_in: _img_val, x_label: _label_val}
_, _accuracy,_accuracy_adver, _prob_max,_prob_adver_max,_label_val = sess.run([logits_cl,\
accuracy,\
accuracy_adver,\
prob_max,\
prob_adver_max,\
label_val],\
feed_dict=feed_vars)
accuracy_list.append(_accuracy)
accuracy_list.append(_accuracy_adver)
print(
'setp:%d, test, (accuracy_A, accuracy_B) = (%.3f, %.3f)' %
(i, _accuracy, _accuracy_adver))
#print('step:{}, test_real_label is:{}'.format(i,_label_val) )
#print('step:{}, test_prob_label is:{}'.format(i, _prob_max) )
coord.request_stop()
coord.join(threads)
sess.close()
aver_accuracy = sum(accuracy_list) / len(accuracy_list)
print('At last, test, aver_accuracy:%.4f' % (aver_accuracy))
print('time use is %d second' % (time.time() - start_time))
return aver_accuracy
def run_training(number_of_classes = 5,\
batch_size=10,\
learning_rate=0.00001,\
num_train_img=90,\
num_epoc=200,\
hide_prob=0.1,\
sigma = 0.6,\
dataset_dir=None,\
logs_train_dir=None,\
checkpoint_dir=None,\
checkpoint_exclude_scopes=None):
if not os.path.exists(logs_train_dir):
os.makedirs(logs_train_dir)
#定义输入变量
tf.reset_default_graph()
with tf.Graph().as_default() as g:
# 生成训练数据
img_train, label_train = gen_data_batch(dataset_dir=dataset_dir,
batch_size=batch_size,
Train=True)
# 推理、训练、准确度
#logits, end_points = inference(x_img, number_of_classes,True)
x_img_in = tf.placeholder(tf.float32, [batch_size, 224, 224, 3])
x_label = tf.placeholder(tf.int64, [batch_size])
label_hot = tf.one_hot(x_label, number_of_classes)
label_advers = number_of_classes - 1 - x_label
label_advers_hot = tf.one_hot(label_advers, number_of_classes)
layer_name1 = 'resnet_v2_152/block4'
target_conv_layer,target_grad_ac,target_grad_yc,logits_cl,end_points = forward_tran(x_img=x_img_in,\
label_index=x_label,\
number_of_classes=number_of_classes,\
layer_name=layer_name1,\
Training=True)
conv_in = end_points[layer_name1]
mask_conv_imgs,_=batch_mask_images(batch_size=batch_size,\
img_input = None,\
img_conv = conv_in,\
target_conv_layer=target_conv_layer,\
target_grad_ac=target_grad_ac,\
target_grad_yc=target_grad_yc,\
sigma=sigma)
_,_,_,logits_adver,_ = forward_tran_advers(x_img = mask_conv_imgs,\
label_index = label_advers,\
number_of_classes = number_of_classes,\
Training = True)
total_loss,logit_cl_loss,logit_adver_loss,L2_loss = losses(logits_cl,\
logits_adver,\
label_hot,\
label_advers_hot,\
number_of_classes)
train_op, global_step = trainning(total_loss, learning_rate)
train_accuracy = (evaluation(logits_cl, label_hot) +
evaluation(logits_adver, label_advers_hot)) / 2.0
#
config = tf.ConfigProto() # 配置GPU参数
config.gpu_options.allow_growth = True # 动态分配GPU资源
#config.gpu_options.per_process_gpu_memory_fraction = 0.85 # 占用GPU90%的显存
sess = tf.Session(config=config)
#
summary_op = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(logs_train_dir, sess.graph)
start_time = time.time()
with sess: #开始一个会话
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
saver = tf.train.Saver()
exclusions = []
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
break
if not excluded:
variables_to_restore.append(var)
print('*******run_trainging11*******')
ckpt = tf.train.get_checkpoint_state(logs_train_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('checkpoint restored from [{0}]'.format(
ckpt.model_checkpoint_path))
print('model_checkpoint_path Loading success')
elif checkpoint_dir:
read_weights_restore = slim.assign_from_checkpoint_fn(
checkpoint_dir, variables_to_restore)
read_weights_restore(sess)
print('checkpoint restored from [{0}]'.format(checkpoint_dir))
else:
print('No checkpoint before training')
pass
print('******run_training12******')
one_epock_step = num_train_img // batch_size
MAX_STEP = num_epoc * one_epock_step
for step in range(num_epoc * num_train_img // batch_size):
#for step in range(1):
_img_train, _label_train = sess.run([img_train, label_train])
patch_img = patch_epock_img(img= _img_train,step=step,epock_step=one_epock_step,\
hide_prob=hide_prob)
feed_vars = {x_img_in: patch_img, x_label: _label_train}
_global_step, _ , _total_loss,_logit_cl_loss, _logit_adver_loss,\
_L2_loss,_accuracy,summary_str = sess.run([global_step,train_op,total_loss,\
logit_cl_loss,logit_adver_loss,\
L2_loss,train_accuracy,summary_op],\
feed_dict=feed_vars)
#每迭代50次，打印出一次结果
if step % 100 == 0:
print('Step %d, total_loss = %.2f, cl_loss=%.2f, adver_loss=%.2f, L2_loss=%.2f, accuracy = %.3f'\
%(_global_step, _total_loss,_logit_cl_loss, _logit_adver_loss, _L2_loss, _accuracy ))
train_writer.add_summary(summary_str, _global_step)
#每迭代700次，利用saver.save()保存一次模型文件，以便测试的时候使用
if ((step % 1200 == 0) and step > 0) or (step + 1) == MAX_STEP:
checkpoint_path = os.path.join(logs_train_dir,
'model_Resnet152.ckpt')
saver.save(sess, checkpoint_path)
print('%.2f sec/step' % ((time.time() - start_time) /
(step + 1e-5)))
coord.request_stop()
coord.join(threads)
sess.close()
print('time use is %d second' % (time.time() - start_time))