Dataset&TFrecord编写CNN代码，loss不下降，预测概率相当于随机...

查看全部 · 发表于 2019-1-17 15:46:35

快来加入 TensorFlowers 大家庭！

您需要登录才可以下载或查看，没有帐号？加入社区

x

数据有4类，都是图片，搭建的网络结构在matlab中表现很好。但在tensorflow中，尝试了很多参数，梯度算法，结果都是loss很高，或者一会大一会小。预测概率永远是0.25（总共有4类），下面是代码，求解答，是我哪里写错了吗

import os
import tensorflow as tf
import numpy as np
import tensorflow.contrib.slim as slim
def lenet_slim(x,is_tra):
with slim.arg_scope([slim.conv2d,slim.fully_connected],
biases_initializer=tf.constant_initializer(0.1),
weights_regularizer=slim.l2_regularizer(0.004),
activation_fn=tf.nn.relu,
weights_initializer=tf.truncated_normal_initializer(stddev=0.1)
):
x=slim.conv2d(x,30,[5,5],scope='conv1',reuse=tf.AUTO_REUSE)
x=slim.batch_norm(x,is_training=is_tra,scope='BN1',reuse=tf.AUTO_REUSE)
x=slim.max_pool2d(x,2,stride=2,scope='pool1')
# print(x)
x=slim.conv2d(x,50,[5,5],scope='conv2',reuse=tf.AUTO_REUSE)
x=slim.batch_norm(x,is_training=is_tra,scope='BN2',reuse=tf.AUTO_REUSE)
x=slim.max_pool2d(x,2,stride=2,scope='pool2')
# x=slim.conv2d(x,64,[3,3],scope='conv3',reuse=tf.AUTO_REUSE)
# x=slim.batch_norm(x,is_training=is_tra,scope='BN3',reuse=tf.AUTO_REUSE)
# x=slim.max_pool2d(x,2,stride=2,scope='pool3')
x=slim.flatten(x,scope="flatten")
# print(x)
# x=slim.fully_connected(x,512,scope='fc1')
x=slim.fully_connected(x,400,scope='fc2',reuse=tf.AUTO_REUSE)
# x=slim.dropout(x,0.5,is_training=is_tra,scope="drop1")
# x=slim.fully_connected(x,128,scope='fc3',reuse=tf.AUTO_REUSE)
# print(x)
logits=slim.fully_connected(x,4,scope='finally',reuse=tf.AUTO_REUSE)
return logits
tf.reset_default_graph()
def parser(record):
features=tf.parse_single_example(record,
features={ 'image_raw':tf.FixedLenFeature([],tf.string),
'label':tf.FixedLenFeature([],tf.int64)})
decode_image=tf.decode_raw(features['image_raw'],tf.uint8)
# decode_image.set_shape([Image_Size,Image_Size,Num_Channels])
decode_image=tf.reshape(decode_image,[256,256,3])
decode_image=tf.cast(decode_image,tf.float32)
label=features['label']
label=tf.cast(label,tf.float32)
return decode_image,label
# train_file_path='D:\\Data\\Data_Picture_useML\\tensorflow_format\\UC_4Class_train.tfrecords'
train_file_path=tf.placeholder(tf.string)
dataset=tf.data.TFRecordDataset(train_file_path)
dataset=dataset.map(parser)
dataset=dataset.shuffle(100).batch(32)
dataset=dataset.repeat(3)
iterator=dataset.make_initializable_iterator()
f1,f2=iterator.get_next()
logits=lenet_slim(f1,True)
global_step=tf.Variable(0,trainable=False)
# prediction=tf.argmax(logits,1)
# prediction=tf.cast(prediction,tf.int32)
f2=tf.cast(f2,tf.int32)
# print([prediction,f2])
# cross_entropy=tf.nn.softmax_cross_entropy_with_logits_v2(logits=prediction,labels=f2)
cross_entropy=tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,labels=f2)
# cross_entropy=tf.nn.softmax_cross_entropy_with_logits(logits=prediction,labels=f2)
cross_mean=tf.reduce_mean(cross_entropy)
Learn_Rate=0.001
Learn_Decay=0.98
learn_rate=tf.train.exponential_decay(Learn_Rate,global_step,4,Learn_Decay,staircase=True)
train_step=tf.train.MomentumOptimizer(learn_rate,momentum=0.7).minimize(cross_mean)
# train_step=tf.train.GradientDescentOptimizer(0.01).minimize(cross_mean,global_step=global_step)
#测试数据集
test_file_path=tf.placeholder(tf.string)
test_dataset=tf.data.TFRecordDataset(test_file_path)
test_dataset=test_dataset.map(parser)
test_dataset=test_dataset.shuffle(100).batch(32)
test_iter=test_dataset.make_initializable_iterator()
test_image_batch,test_label_batch=test_iter.get_next()
# #测试成功率
test_logit=lenet_slim(test_image_batch,False)
prediction=tf.argmax(test_logit,axis=-1,output_type=tf.int32)
with tf.Session() as sess:
# for i in range(10):
# f3,f4=sess.run([f1,f2])
# print(f3)
#一定要加tf.local_variables_initializer()，不然
sess.run((tf.global_variables_initializer(),tf.local_variables_initializer()))
sess.run(iterator.initializer,feed_dict={train_file_path:'D:\\Data\\Data_Picture_useML\\tensorflow_format\\UC_4Class_train.tfrecords'})
while True:
try:
# f3,f4=sess.run([f1,f2])
f5,_=sess.run([cross_mean,train_step])
# f7=sess.run(cross_entropy)
# print(f7)
print(f5)
# print(f4.shape)
except tf.errors.OutOfRangeError:
break
print('train finish')
sess.run(test_iter.initializer,feed_dict={test_file_path:'D:\\Data\\Data_Picture_useML\\tensorflow_format\\UC_4Class_test.tfrecords'})
test_results=[]
test_labels=[]
while True:
try:
# t3,t4=sess.run([test_logit,prediction])
# print([t3[0:2],t4[0:2]])
log,pred,label=sess.run([test_logit,prediction,test_label_batch])
print([log[0:2],pred[0:2]])
test_results.extend(pred)
test_labels.extend(label)
# t1,t2=sess.run([test_image_batch,test_label_batch])
# print(t2.shape)
# print("/n")
except tf.errors.OutOfRangeError:
break
# 计算准确率
print('11',test_results,test_labels)
correct = [float(y == y_) for (y, y_) in zip (test_results, test_labels)]
print("准确率:",correct)
accuracy = sum(correct) / len(correct)
print("Test accuracy is:", accuracy)

复制代码

我知道答案回答被采纳将会获得10 金币 + 20 金币已有0人回答

Dataset&TFrecord编写CNN代码，loss不下降，预测概率相当于随机...

快来加入 TensorFlowers 大家庭！

TF 使用案例

新帖排行

热点动态