def test_network():
if not ct.TEST_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=100,
validation_percentage=0,
inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,
tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
test_image_num = int(dataSetSizeList['testing'])
image_inputs = tf.placeholder(
tf.float32, (1, ct.INPUT_SIZE, ct.INPUT_SIZE, ct.IMAGE_CHANNEL * 2),
'testing_inputs')
label_inputs = tf.placeholder(tf.float32, (1, ct.CLASS_NUM),
'testing_outputs')
nn_output = foward_propagation(image_inputs, is_training=False)
correct_prediction = tf.equal(tf.argmax(nn_output, 1),
tf.argmax(label_inputs, 1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor, label_tensor = readImageFromTFRecord(
ct.CATELOGS[1], tfrecord_dir=ct.TEST_TFRECORD_DIR)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
correct_prediction_list = []
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
for _ in range(test_image_num):
# for _ in range(800):
test_image, test_label = sess.run([image_tensor, label_tensor])
per_correct_prediction = sess.run(correct_prediction,
feed_dict={
image_inputs:
[test_image],
label_inputs:
[test_label]
})
correct_prediction_list.append(per_correct_prediction[0])
correct_num = 0
for rst in correct_prediction_list:
correct_num += rst
accuracy_score = correct_num / len(correct_prediction_list)
print('after %s iteration, the testing accuracy is %g' %
(global_step, accuracy_score))
else:
print('no model')
coord.request_stop()
coord.join(threads)
def validate_network():
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs = tf.placeholder(
tf.float32, (1, ct.INPUT_SIZE, ct.INPUT_SIZE, ct.IMAGE_CHANNEL * 2),
'validation_inputs')
label_inputs = tf.placeholder(tf.float32, (1, ct.CLASS_NUM),
'validation_outputs')
nn_output = foward_propagation(image_inputs)
correct_prediction = tf.equal(tf.argmax(nn_output, 1),
tf.argmax(label_inputs, 1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor, label_tensor = readImageFromTFRecord(ct.CATELOGS[2])
saver = tf.train.Saver()
with tf.Session() as sess:
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while (True):
correct_prediction_list = []
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
for _ in range(validation_image_num):
test_image, test_label = sess.run(
[image_tensor, label_tensor])
per_correct_prediction = sess.run(correct_prediction,
feed_dict={
image_inputs:
[test_image],
label_inputs:
[test_label]
})
correct_prediction_list.append(per_correct_prediction[0])
correct_num = 0
for rst in correct_prediction_list:
correct_num += rst
accuracy_score = correct_num / len(correct_prediction_list)
print('after %s iteration, the validation accuracy is %g' %
(global_step, accuracy_score))
else:
print('no model')
print('running..........')
time.sleep(300)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,validation_percentage=100,inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
# dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs=tf.placeholder(tf.float32, (1,None,None,ct.IMAGE_CHANNEL*2), 'validation_inputs')
# image_inputs = tf.cond(tf.equal(input_shape_flag,0),lambda:image_inputs_256_256,
# lambda:tf.cond(tf.equal(input_shape_flag,1),lambda:image_inputs_181_362,lambda:image_inputs_362_181))
# image_inputs=tf.placeholder(tf.float32, (1,input_size[0],input_size[1],ct.IMAGE_CHANNEL*2), 'validation_inputs')
label_inputs =tf.placeholder(tf.float32,(1,ct.CLASS_NUM), 'validation_outputs')
nn_output,layer_outputs_tensor = forward_propagation(image_inputs,is_training=False)
label_value_tensor = tf.argmax(label_inputs,1)
pred_value_tensor = tf.argmax(nn_output,1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor,label_tensor,proportion_tensor= readImageFromTFRecord(ct.CATELOGS[2],tfrecord_dir= ct.TEST_TFRECORD_DIR,num_epochs=None)
saver = tf.train.Saver()
with tf.Session() as sess :
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while(True):
TP = 0
FN = 0
FP = 0
TN = 0
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess,ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
if int(global_step)>ct.STEPS+1:
break
for i in range(validation_image_num):
test_image, test_label, proportion= sess.run([image_tensor,label_tensor,proportion_tensor])
result=[0,0,0]
prediction = 0
input_size = ((360,180),(180,360),(256,256))
for j in range(3):
test_image = cv2.resize(test_image,input_size[j],interpolation=cv2.INTER_LINEAR)
pred,label,layer_outputs = sess.run([pred_value_tensor,label_value_tensor,layer_outputs_tensor], feed_dict= {image_inputs:[test_image],label_inputs:[test_label]})
result[j]=pred[0];
if sum(result)>=2:
prediction = 1
if label[0]:
if prediction:
TP+=1
else:
FN+=1
else:
if not prediction:
TN+=1
else:
FP+=1
accuracy = (TP+TN)/(TP+FN+TN+FP)
precision = TP/(TP+FP+1e-8)
recall = TP/(TP+FN)
f1 = 2*precision*recall/(precision+recall+1e-8)
print('after %s iteration, the accuracy is %g,precision is %g,recall is %g,F1 is %g'%(global_step,accuracy,precision,recall,f1))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
print('running..........')
time.sleep(100)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,validation_percentage=100,inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
# dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs=tf.placeholder(tf.float32, (1,None,None,ct.IMAGE_CHANNEL*2), 'validation_inputs')
# image_inputs = tf.cond(tf.equal(input_shape_flag,0),lambda:image_inputs_256_256,
# lambda:tf.cond(tf.equal(input_shape_flag,1),lambda:image_inputs_181_362,lambda:image_inputs_362_181))
# image_inputs=tf.placeholder(tf.float32, (1,input_size[0],input_size[1],ct.IMAGE_CHANNEL*2), 'validation_inputs')
label_inputs =tf.placeholder(tf.float32,(1,ct.CLASS_NUM), 'validation_outputs')
nn_output,layer_outputs_tensor = forward_propagation(image_inputs,is_training=False)
label_value_tensor = tf.argmax(label_inputs,1)
pred_value_tensor = tf.argmax(nn_output,1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor,label_tensor,proportion_tensor= readImageFromTFRecord(ct.CATELOGS[2],tfrecord_dir= ct.TEST_TFRECORD_DIR,num_epochs=None)
saver = tf.train.Saver()
with tf.Session() as sess :
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while(True):
TP = 0
FN = 0
FP = 0
TN = 0
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess,ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
if int(global_step)>ct.STEPS+1:
break
for i in range(validation_image_num):
test_image, test_label, proportion= sess.run([image_tensor,label_tensor,proportion_tensor])
if proportion<1/1.5:
input_size = (360,180)
elif proportion >1.5:
# input_shape_flag = 2
input_size = (180,360)
else:
# input_shape_flag = 0
input_size = (256,256)
# feature_map_path=os.path.join(ct.FEATURE_MAP,str(i))
# isFileExist(feature_map_path)
test_image = cv2.resize(test_image,input_size,interpolation=cv2.INTER_LINEAR)
# a,b = cv2.split(test_image)
# cv2.imwrite(os.path.join(feature_map_path,'hist.jpg'),a*255)
# cv2.imwrite(os.path.join(feature_map_path,'curr.jpg'),b*255)
# cv2.namedWindow('1',0)
# cv2.namedWindow('2',0)
# cv2.imshow('1',a)
# cv2.imshow('2',b)
# cv2.waitKey()
pred,label,layer_outputs = sess.run([pred_value_tensor,label_value_tensor,layer_outputs_tensor], feed_dict= {image_inputs:[test_image],label_inputs:[test_label]})
#
# feature_map = layer_outputs['resnet_v2_50/conv1']
# predict = layer_outputs['prediction']
# with open(os.path.join(feature_map_path,'predict'),'w') as f:
# f.write(str(predict))
# f.write('\n')
#
# feature_map = np.squeeze(feature_map)
# feature_map = cv2.split(feature_map)
# # cv2.namedWindow('3',0)
# for i in range(len(feature_map)):
# cv2.imwrite(os.path.join(feature_map_path,str(i)+'.jpg'),feature_map[i]*255)
# # cv2.imshow('3',feature_map[i])
# # cv2.waitKey()
if label[0]:
if pred[0]:
TP+=1
else:
FN+=1
else:
if not pred[0]:
TN+=1
else:
FP+=1
accuracy = (TP+TN)/(TP+FN+TN+FP)
precision = TP/(TP+FP+1e-8)
recall = TP/(TP+FN)
f1 = 2*precision*recall/(precision+recall+1e-8)
print('after %s iteration, the accuracy is %g,precision is %g,recall is %g,F1 is %g'%(global_step,accuracy,precision,recall,f1))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
print('running..........')
time.sleep(200)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,
validation_percentage=100,
inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,
tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
# dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs = tf.placeholder(
tf.float32, (1, ct.INPUT_SIZE, ct.INPUT_SIZE, ct.IMAGE_CHANNEL * 2),
'validation_inputs')
label_inputs = tf.placeholder(tf.float32, (1, ct.CLASS_NUM),
'validation_outputs')
nn_output, _ = foward_propagation(image_inputs, is_training=False)
label_value_tensor = tf.argmax(label_inputs, 1)
pred_value_tensor = tf.argmax(nn_output, 1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor, label_tensor = readImageFromTFRecord(
ct.CATELOGS[2], tfrecord_dir=ct.TEST_TFRECORD_DIR)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while (True):
TP = 0
FN = 0
FP = 0
TN = 0
# sample_num = [0 for _ in range(7)]
# correct_num = [0 for _ in range(7)]
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
for _ in range(validation_image_num):
test_image, test_label = sess.run(
[image_tensor, label_tensor])
pred, label = sess.run(
[pred_value_tensor, label_value_tensor],
feed_dict={
image_inputs: [test_image],
label_inputs: [test_label]
})
if label[0]:
if pred[0]:
TP += 1
else:
FN += 1
else:
if not pred[0]:
TN += 1
else:
FP += 1
# index = label[0]
# sample_num[index]+=1
# if pred[0] == index:
# correct_num[index]+=1
# print(sample_num)
# print(positive_sample_num)
# print(negative_sample_num)
accuracy = (TP + TN) / (TP + FN + TN + FP)
precision = TP / (TP + FP + 1e-8)
recall = TP / (TP + FN)
f1 = 2 * precision * recall / (precision + recall + 1e-8)
# correct_num = sum(correct_num)
# accuracy_score = correct_num / sum(sample_num)
# print('after %s iteration, the validation accuracy is %g'%(global_step,accuracy_score))
print(
'after %s iteration, the accuracy is %g,precision is %g,recall is %g,F1 is %g'
% (global_step, accuracy, precision, recall, f1))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
if int(global_step) > ct.STEPS:
break
print('running..........')
time.sleep(100)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,
validation_percentage=100,
inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,
tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs = tf.placeholder(
tf.float32, (1, ct.INPUT_SIZE, ct.INPUT_SIZE, ct.IMAGE_CHANNEL * 2),
'validation_inputs')
label_inputs = tf.placeholder(tf.float32, (1, ct.CLASS_NUM),
'validation_outputs')
nn_output = foward_propagation(image_inputs, is_training=False)
label_value_tensor = tf.argmax(label_inputs, 1)
pred_value_tensor = tf.argmax(nn_output, 1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor, label_tensor = readImageFromTFRecord(
ct.CATELOGS[2], tfrecord_dir=ct.TEST_TFRECORD_DIR, num_epochs=None)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while (True):
# positive_sample_num=0
# negative_sample_num=0
# p2p = 0
# n2n=0
sample_num = [0 for _ in range(ct.CLASS_NUM)]
correct_num = [0 for _ in range(ct.CLASS_NUM)]
proportion = [0 for _ in range(ct.CLASS_NUM)]
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
if int(global_step) > ct.STEPS + 1:
break
for _ in range(validation_image_num):
test_image, test_label = sess.run(
[image_tensor, label_tensor])
pred, label = sess.run(
[pred_value_tensor, label_value_tensor],
feed_dict={
image_inputs: [test_image],
label_inputs: [test_label]
})
# if label[0]:
# positive_sample_num+=1
# if pred[0]:
# p2p+=1
# else:
# negative_sample_num+=1
# if not pred[0]:
# n2n+=1
index = label[0]
sample_num[index] += 1
if pred[0] == index:
correct_num[index] += 1
# print(sample_num)
# print(positive_sample_num)
# print(negative_sample_num)
# correct_num = p2p+ n2n
# accuracy_score = correct_num/(positive_sample_num+negative_sample_num)
# p2p_score = p2p/positive_sample_num
# n2n_score = n2n/negative_sample_num
accuracy_score = sum(correct_num) / sum(sample_num)
proportion = [
correct_num[i] / sample_num[i] for i in range(ct.CLASS_NUM)
]
print('after %s iteration, the validation accuracy is %g' %
(global_step, accuracy_score))
print(
'stain:%g, luminance:%g, rotation:%g, abnormal:%g, foreignBody:%g, character:%g'
% (proportion[0], proportion[1], proportion[2],
proportion[3], proportion[4], proportion[5]))
# print('after %s iteration, the validation accuracy is %g,p2p_score is %g,n2n_score is %g'%(global_step,accuracy_score,p2p_score,n2n_score))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
print('running..........')
time.sleep(100)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,validation_percentage=100,inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
# dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
# image_inputs=tf.placeholder(tf.float32, (1,ct.INPUT_SIZE,ct.INPUT_SIZE,ct.IMAGE_CHANNEL*2), 'validation_inputs')
# label_inputs =tf.placeholder(tf.float32,(1,ct.CLASS_NUM), 'validation_outputs')
curr_image_inputs=tf.placeholder(tf.float32, (1,ct.INPUT_SIZE,ct.INPUT_SIZE,ct.IMAGE_CHANNEL), 'curr_inputs')
hist_image_inputs=tf.placeholder(tf.float32, (1,ct.INPUT_SIZE,ct.INPUT_SIZE,ct.IMAGE_CHANNEL), 'hist_inputs')
label_inputs =tf.placeholder(tf.float32,(1,ct.CLASS_NUM), 'outputs')
nn_output,_ = forward_propagation(curr_image_inputs,hist_image_inputs,is_training=False)
label_value_tensor = tf.argmax(label_inputs,1)
pred_value_tensor = tf.argmax(nn_output,1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
curr_img_tensor,hist_img_tensor,label_tensor= readImageFromTFRecord(ct.CATELOGS[2],tfrecord_dir= ct.TEST_TFRECORD_DIR)
# curr_img_tensor=tf.reshape(curr_img_tensor,[1,ct.INPUT_SIZE,ct.INPUT_SIZE,ct.IMAGE_CHANNEL])
# hist_img_tensor = tf.reshape(hist_img_tensor,[1,ct.INPUT_SIZE,ct.INPUT_SIZE,ct.IMAGE_CHANNEL])
# label_tensor = tf.reshape(label_tensor,[1,ct.CLASS_NUM])
saver = tf.train.Saver()
with tf.Session() as sess :
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while(True):
sample_num = [0 for _ in range(7)]
correct_num = [0 for _ in range(7)]
inference_time=0
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess,ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
for _ in range(validation_image_num):
# start_time = time.time()
curr_img,hist_img,label = sess.run([curr_img_tensor,hist_img_tensor,label_tensor])
pred,label = sess.run([pred_value_tensor,label_value_tensor], feed_dict= {curr_image_inputs:[curr_img],hist_image_inputs:[hist_img],label_inputs:[label]})
# inference_time+=(time.time()-start_time)
index = label[0]
sample_num[index]+=1
if pred[0] == index:
correct_num[index]+=1
# print(sample_num)
# print(positive_sample_num)
# print(negative_sample_num)
# accuracy = (TP+TN)/(TP+FN+TN+FP)
# precision = TP/(TP+FP+1e-8)
# recall = TP/(TP+FN)
# f1 = 2*precision*recall/(precision+recall+1e-8)
# print(inference_time/(TP+FN+TN+FP))
accuracy_score = sum(correct_num) / sum(sample_num)
proportion = [correct_num[i]/sample_num[i] for i in range(ct.CLASS_NUM)]
print('after %s iteration, the validation accuracy is %g'%(global_step,accuracy_score))
print('stain:%g, luminance:%g, rotation:%g, abnormal:%g, foreignBody:%g, character:%g'%(proportion[0],proportion[1],proportion[2],proportion[3],proportion[4],proportion[5]))
# print('after %s iteration, the accuracy is %g,precision is %g,recall is %g,F1 is %g'%(global_step,accuracy,precision,recall,f1))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
if int(global_step)>ct.STEPS:
break
print('running..........')
time.sleep(100)
coord.request_stop()
coord.join(threads)
def validate_network():
if not ct.VALIDATION_PERCENTAGE:
loadImageAndConvertToTFRecord(test_percentage=0,
validation_percentage=100,
inputDataDir=ct.TEST_DATASET_PATH,
infoSavePath=ct.TEST_INFOMATION_PATH,
tfrecordPath=ct.TEST_TFRECORD_DIR)
dataSetSizeList = readInfoFromFile(ct.TEST_INFOMATION_PATH)
else:
dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
# dataSetSizeList = readInfoFromFile(ct.INFORMATION_PATH)
validation_image_num = int(dataSetSizeList['validation'])
image_inputs = tf.placeholder(
tf.float32, (1, ct.INPUT_SIZE, ct.INPUT_SIZE, ct.IMAGE_CHANNEL * 2),
'validation_inputs')
label_inputs = tf.placeholder(tf.float32, (1, ct.CLASS_NUM),
'validation_outputs')
nn_output, layer_outputs_tensor = foward_propagation(image_inputs,
is_training=False)
label_value_tensor = tf.argmax(label_inputs, 1)
pred_value_tensor = tf.argmax(nn_output, 1)
# correct_prediction = tf.equal(tf.argmax(nn_output,1), tf.argmax(label_inputs,1))
# accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
image_tensor, label_tensor = readImageFromTFRecord(
ct.CATELOGS[2], tfrecord_dir=ct.TEST_TFRECORD_DIR)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.local_variables_initializer().run()
tf.global_variables_initializer().run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
while (True):
# TP = 0
# FN = 0
# FP = 0
# TN = 0
# sample_num = [0 for _ in range(7)]
# correct_num = [0 for _ in range(7)]
ckpt = tf.train.get_checkpoint_state(ct.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
for i in range(validation_image_num):
test_image, test_label = sess.run(
[image_tensor, label_tensor])
pred, label, layer_outputs = sess.run([
pred_value_tensor, label_value_tensor,
layer_outputs_tensor
],
feed_dict={
image_inputs:
[test_image],
label_inputs:
[test_label]
})
feature_map_path = os.path.join(ct.FEATURE_MAP, str(i))
isFileExist(feature_map_path)
feature_map = layer_outputs[
'resnet_v2_50/block3/unit2/bottleneck_v2']
predict = layer_outputs['prediction']
with open(os.path.join(feature_map_path, 'predict'),
'w') as f:
f.write(str(predict))
f.write('\n')
feature_map = np.squeeze(feature_map)
feature_map = cv2.split(feature_map)
# cv2.namedWindow('3',0)
for i in range(len(feature_map)):
cv2.imwrite(
os.path.join(feature_map_path,
str(i) + '.jpg'),
feature_map[i] * 255)
# cv2.imshow('3',feature_map[i])
# cv2.waitKey()
# pred,label = sess.run([pred_value_tensor,label_value_tensor], feed_dict= {image_inputs:[test_image],label_inputs:[test_label]})
# if label[0]:
# if pred[0]:
# TP+=1
# else:
# FN+=1
# else:
# if not pred[0]:
# TN+=1
# else:
# FP+=1
# index = label[0]
# sample_num[index]+=1
# if pred[0] == index:
# correct_num[index]+=1
# print(sample_num)
# print(positive_sample_num)
# print(negative_sample_num)
# accuracy = (TP+TN)/(TP+FN+TN+FP)
# precision = TP/(TP+FP+1e-8)
# recall = TP/(TP+FN)
# f1 = 2*precision*recall/(precision+recall+1e-8)
# correct_num = sum(correct_num)
# accuracy_score = correct_num / sum(sample_num)
# print('after %s iteration, the validation accuracy is %g'%(global_step,accuracy_score))
# print('after %s iteration, the accuracy is %g,precision is %g,recall is %g,F1 is %g'%(global_step,accuracy,precision,recall,f1))
else:
print('no model')
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# print(sess.run(update_ops))
if int(global_step) > ct.STEPS:
break
print('running..........')
time.sleep(100)
coord.request_stop()
coord.join(threads)