TensorFlow详解猫狗识别（三）--训练过程

Dear 丶 2022-05-14 08:37 273阅读 0赞

## 感悟 ##

在这段时间中，同时测试了几个神经网络的模型（LeNet、AlexNet、VGG16...）

感受到了调节超参数的重要性，单单对于LeNet来说，得出，当：

batch\_size = 32

lr = 0.0001

max\_step = 6000~10000

时函数收敛比较快，训练步数介于6000~10000时，训练出来的效果比较好，之前训练了一晚上100000步，第二天过来看预测结果，满心欢喜看到train\_acc达到了100%，但用来识别精度大概在75%，模型出现了严重的过拟合，训练8000步的时候，预测结果很理想，100张网图，基本没有错。

而当lr设太小的话，模型不收敛或收敛的太慢，lr太大的话模型会出现震荡。

下面直接给出训练代码，里面会有详细注释。

## 代码 ##

import os
    import numpy as np
    import tensorflow as tf
    import test
    import model
    import time
    
    
    N_CLASSES = 2
    IMG_W = 208
    IMG_H = 208
    BATCH_SIZE = 16
    CAPACITY = 2000 #队列中元素个数
    MAX_STEP = 8000
    learning_rate = 0.0001 #小于0.001
    
    print("I'm OK")
    train_dir = 'E:\\Pycharm\\tf-01\\Bigwork\\train\\'  # 训练图片文件夹
    logs_train_dir = 'E:\\Pycharm\\tf-01\\Bigwork\\savenet02\\'  # 保存训练结果文件夹
    
    train, train_label = test.get_files(train_dir)
    
    train_batch, train_label_batch = test.get_batch(train,
                                                           train_label,
                                                           IMG_W,
                                                           IMG_H,
                                                           BATCH_SIZE,
                                                           CAPACITY)
    
    
    
    #训练操作定义
    sess = tf.Session()
    
    train_logits = model.inference(train_batch, BATCH_SIZE, N_CLASSES)
    train_loss = model.losses(train_logits, train_label_batch)
    train_op = model.trainning(train_loss, learning_rate)
    train_acc = model.evaluation(train_logits, train_label_batch)
    
    #train_label_batch = tf.one_hot(train_label_batch,2,1,0)
    #测试操作定义
    
    
    summary_op = tf.summary.merge_all()
    
    #产生一个writer来写log文件
    train_writer = tf.summary.FileWriter(logs_train_dir,sess.graph)
    saver = tf.train.Saver()
    
    sess.run(tf.global_variables_initializer())
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess,coord = coord)#加入队列，很重要
    
    tra_loss = .0
    tra_acc = .0
    # val_loss = .0
    # val_acc = .0
    
    try:
        start = time.clock()#计算每一个step所花的时间
        for step in np.arange(MAX_STEP):
            if coord.should_stop():
                break
            _,tra_loss_,tra_acc_ = sess.run([train_op,train_loss,train_acc])
            # val_loss_, val_acc_ = sess.run([test_loss, test_acc])
            #下面这一段为我为了打印神经网络最后一层变化写的，可以不要
            '''
            train,label = sess.run([train_logits,train_label_batch])
            #print(train)
            L = []
            for i in train:
                max_ = np.argmax(i)
                L.append(max_)
            print(L)
            print(label)
            '''
            tra_loss = tra_loss+tra_loss_
            tra_acc = tra_acc+tra_acc_
            # val_loss = val_loss+val_loss_
            # val_acc = val_acc+val_acc_
    
            if (step+1) % 50 == 0 and step!=0:
                end = time.clock()
                print('Step %d, train loss = %.2f, train accuracy = %.2f%%' % (step+1, tra_loss/50, tra_acc * 100.0/50))
                #print('Step %d, val loss = %.2f, val accuracy = %.2f%%' % (step, val_loss/50,val_acc*100.0/50))
                print(str(end-start))
                tra_loss = .0
                tra_acc = .0
                summary_str = sess.run(summary_op)
                train_writer.add_summary(summary_str, step)
    
                start = time.clock()
    
    
            # 每隔2000步，保存一次训练好的模型
            if step%2000==0 or step == MAX_STEP-1:
                checkpoint_path = os.path.join(logs_train_dir, 'model.ckpt')
                saver.save(sess, checkpoint_path, global_step=step)
    
                
    except tf.errors.OutOfRangeError:
        print('Done training -- epoch limit reached')
    
    finally:
        coord.request_stop()
    
    coord.join(threads)
    sess.close()

## 开始训练 ##

下图我是每间隔10个step打印一次结果，可以看到，训练到2000多步的时候，精度已经可以达到了84%。

![70][]

下一篇：[预测][Link 1]

[70]: /images/20220514/dae3aecf2e4f487bbb8d3ee54e8f1792.png
[Link 1]: https://blog.csdn.net/qq_41004007/article/details/82425882