VGG 模型的输出在训练后变得恒定，并且损失/准确率没有提高答案

【问题标题】：Output of VGG Model becomes constant after training and Loss/Accuracy are not improvingVGG 模型的输出在训练后变得恒定，并且损失/准确率没有提高
【发布时间】：2021-05-25 07:41:31
【问题描述】：

我正在尝试实现一个稍微小一点的 VGG16 版本，并在大约 6000 张图像的数据集上从头开始训练它（5400 张用于训练，600 张用于验证）。我选择了 30 的批量大小，以便它可以整齐地放入数据集中，否则我会在训练期间得到他的 IncompatibleShape 错误。

经过 15-20 个 epoch 后，EarlyStopping 回调启动并停止训练。

这个模型我面临两个问题

之后，当我将测试图像传入模型时，输出似乎保持不变。我最不希望的是对于imageA，预测输出应该与 imageB 不同。我无法弄清楚为什么会这样
损失和准确性似乎没有太大变化。我预计对于 epoch 的数量，准确率至少会提高到 50% 左右，但不会超过 23%。我尝试包含 steps_per_epoch、ReduceLROnPlateau，但它们似乎没有任何影响。

训练输出：

Epoch 1/50
180/180 [==============================] - 50s 278ms/step - loss: 1.6095 - categorical_accuracy: 0.1987 - val_loss: 1.6109 - val_categorical_accuracy: 0.1267

Epoch 00001: val_loss improved from inf to 1.61094, saving model to vgg16.h5
Epoch 2/50
180/180 [==============================] - 51s 285ms/step - loss: 1.6095 - categorical_accuracy: 0.2044 - val_loss: 1.6107 - val_categorical_accuracy: 0.2133

Epoch 00002: val_loss improved from 1.61094 to 1.61067, saving model to vgg16.h5
Epoch 3/50
180/180 [==============================] - 51s 285ms/step - loss: 1.6098 - categorical_accuracy: 0.1946 - val_loss: 1.6106 - val_categorical_accuracy: 0.1400

Epoch 00003: val_loss improved from 1.61067 to 1.61059, saving model to vgg16.h5
Epoch 4/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.1928 - val_loss: 1.6098 - val_categorical_accuracy: 0.2000

Epoch 00004: val_loss improved from 1.61059 to 1.60983, saving model to vgg16.h5

Epoch 00004: ReduceLROnPlateau reducing learning rate to 2.5000001187436283e-05.
Epoch 5/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6093 - categorical_accuracy: 0.2033 - val_loss: 1.6103 - val_categorical_accuracy: 0.1467

Epoch 00005: val_loss did not improve from 1.60983
Epoch 6/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6094 - categorical_accuracy: 0.1989 - val_loss: 1.6106 - val_categorical_accuracy: 0.1400

Epoch 00006: val_loss did not improve from 1.60983
Epoch 7/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6094 - categorical_accuracy: 0.2069 - val_loss: 1.6098 - val_categorical_accuracy: 0.1733

Epoch 00007: val_loss improved from 1.60983 to 1.60978, saving model to vgg16.h5

Epoch 00007: ReduceLROnPlateau reducing learning rate to 1e-05.
Epoch 8/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6093 - categorical_accuracy: 0.2076 - val_loss: 1.6103 - val_categorical_accuracy: 0.1600

Epoch 00008: val_loss did not improve from 1.60978
Epoch 9/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.2006 - val_loss: 1.6097 - val_categorical_accuracy: 0.2200

Epoch 00009: val_loss improved from 1.60978 to 1.60975, saving model to vgg16.h5
Epoch 10/50
180/180 [==============================] - 52s 287ms/step - loss: 1.6095 - categorical_accuracy: 0.2043 - val_loss: 1.6101 - val_categorical_accuracy: 0.1667

Epoch 00010: val_loss did not improve from 1.60975
Epoch 11/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6094 - categorical_accuracy: 0.2009 - val_loss: 1.6102 - val_categorical_accuracy: 0.1800

Epoch 00011: val_loss did not improve from 1.60975
Epoch 12/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.2041 - val_loss: 1.6115 - val_categorical_accuracy: 0.1600

Epoch 00012: val_loss did not improve from 1.60975
Epoch 13/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.1989 - val_loss: 1.6108 - val_categorical_accuracy: 0.1867

Epoch 00013: val_loss did not improve from 1.60975
Epoch 14/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6094 - categorical_accuracy: 0.2009 - val_loss: 1.6102 - val_categorical_accuracy: 0.1733

Epoch 00014: val_loss did not improve from 1.60975
Epoch 15/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6093 - categorical_accuracy: 0.2074 - val_loss: 1.6113 - val_categorical_accuracy: 0.1467

Epoch 00015: val_loss did not improve from 1.60975
Epoch 16/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6098 - categorical_accuracy: 0.1983 - val_loss: 1.6105 - val_categorical_accuracy: 0.1867

Epoch 00016: val_loss did not improve from 1.60975
Epoch 17/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.2056 - val_loss: 1.6119 - val_categorical_accuracy: 0.1667

Epoch 00017: val_loss did not improve from 1.60975
Epoch 18/50
180/180 [==============================] - 52s 286ms/step - loss: 1.6093 - categorical_accuracy: 0.1994 - val_loss: 1.6110 - val_categorical_accuracy: 0.1800

Epoch 00018: val_loss did not improve from 1.60975
Epoch 19/50
180/180 [==============================] - 51s 286ms/step - loss: 1.6095 - categorical_accuracy: 0.2026 - val_loss: 1.6103 - val_categorical_accuracy: 0.1667

Epoch 00019: val_loss did not improve from 1.60975
Restoring model weights from the end of the best epoch.
Epoch 00019: early stopping

用于获取预测的代码：

predictions = []
actuals=[]

for i, (images, labels) in enumerate( test_datasource):
  if i > 2:
    break
  pred = model_2(images)
  print(labels.shape, pred.shape)
  for j in range(len(labels)):
    actuals.append( labels[j])
    predictions.append(pred[j])
    print(labels[j].numpy(), "\t", pred[j].numpy())

以上代码的输出：

(30, 5) (30, 5)
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
(30, 5) (30, 5)
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
(30, 5) (30, 5)
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 0. 1.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[1. 0. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 1. 0. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 0. 1. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]
[0. 0. 1. 0. 0.]     [0.19907779 0.20320047 0.1968051  0.20173152 0.19918515]

这是模型摘要：

Model: "vgg16"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         [(30, 224, 224, 3)]       0         
_________________________________________________________________
conv_1_1 (Conv2D)            (30, 224, 224, 32)        896       
_________________________________________________________________
conv_1_2 (Conv2D)            (30, 224, 224, 32)        9248      
_________________________________________________________________
maxpool_1 (MaxPooling2D)     (30, 112, 112, 32)        0         
_________________________________________________________________
conv_2_1 (Conv2D)            (30, 112, 112, 64)        18496     
_________________________________________________________________
conv_2_2 (Conv2D)            (30, 112, 112, 64)        36928     
_________________________________________________________________
maxpool_2 (MaxPooling2D)     (30, 56, 56, 64)          0         
_________________________________________________________________
conv_3_1 (Conv2D)            (30, 56, 56, 128)         73856     
_________________________________________________________________
conv_3_2 (Conv2D)            (30, 56, 56, 128)         147584    
_________________________________________________________________
conv_3_3 (Conv2D)            (30, 56, 56, 128)         147584    
_________________________________________________________________
maxpool_3 (MaxPooling2D)     (30, 28, 28, 128)         0         
_________________________________________________________________
conv_4_1 (Conv2D)            (30, 28, 28, 256)         295168    
_________________________________________________________________
conv_4_2 (Conv2D)            (30, 28, 28, 256)         590080    
_________________________________________________________________
conv_4_3 (Conv2D)            (30, 28, 28, 256)         590080    
_________________________________________________________________
maxpool_4 (MaxPooling2D)     (30, 14, 14, 256)         0         
_________________________________________________________________
conv_5_1 (Conv2D)            (30, 14, 14, 256)         590080    
_________________________________________________________________
conv_5_2 (Conv2D)            (30, 14, 14, 256)         590080    
_________________________________________________________________
conv_5_3 (Conv2D)            (30, 14, 14, 256)         590080    
_________________________________________________________________
maxpool_5 (MaxPooling2D)     (30, 7, 7, 256)           0         
_________________________________________________________________
flatten (Flatten)            (30, 12544)               0         
_________________________________________________________________
fc_1 (Dense)                 (30, 4096)                51384320  
_________________________________________________________________
dropout_1 (Dropout)          (30, 4096)                0         
_________________________________________________________________
fc_2 (Dense)                 (30, 4096)                16781312  
_________________________________________________________________
dropout_2 (Dropout)          (30, 4096)                0         
_________________________________________________________________
output (Dense)               (30, 5)                   20485     
=================================================================
Total params: 71,866,277
Trainable params: 71,866,277
Non-trainable params: 0

代码在 Google Colab 中：https://colab.research.google.com/drive/1AWe87Zb3MvF90j3RS7sv3OiSgR86q4j_

我尝试了 VGG-16 的两个版本，一个是滤镜深度的一半，第二个是滤镜深度的四分之一。

【问题讨论】：

似乎即使您的训练损失在每个 epoch 之后都没有改善，而在每个 epoch 之后训练几乎相同，请再次检查训练管道，或者请在此处提及训练管道以供其他人查看
训练管道是指model.summary的输出吗？抱歉，我还在学习该领域的行话
抱歉没有注意到您已经提供了google colab链接
你能不能试着提高你的学习率，并检查你的损失函数是否得到了它所期望的正确输入，tensorflow.org/api_docs/python/tf/keras/losses/…
我在加载数据时使用 label_mode='categorical' 将标签转换为 one-hot 编码向量。标签确实是以这种方式生成的，如上所示，我将标签与预测输出进行比较。 CategoricalCrossEntropy 似乎是 one-hot 编码场景中通常使用的损失。您要求检查损失函数还有其他原因吗？我现在将尝试提高学习率。完成后会添加评论

标签： python tensorflow keras vgg-net image-classification

【解决方案1】：

我的印象是，问题在于损失和准确性是不变的，这就是为什么在使用模型进行预测时，无论输入什么输入，输出都是固定不变的。

当我通过调用 Model(inputs...,outputs...) 函数重新初始化模型并将输入传递给它而不进行训练时，输出至少发生了变化。

我尝试了多种学习率和优化器，但模型的行为没有改变。

经过更多谷歌搜索后，我偶然发现了这些文章：

我对代码进行了两次更改以使其最终工作。

我最初只是将图像数组除以 255 以使值介于 0 和 1 之间，然后从结果中减去 0.5 以使值介于 -0.5 和 0.5 之间。这已更改为使用 tf.image.per_image_standardization(images-127) 并将结果除以每个图像中的最大值。结果，图像值介于 -1 和 +1 之间
固定输出的另一个主要原因是模型的 relu 单元在训练期间死亡（或饱和）。 relu 激活函数本身就有这个问题，一旦变量的权重变为 0，它就无法从中恢复。尽管据说高学习率会导致此问题，但我无法找到缓解此问题的学习率。另一种解决方案是将激活函数更改为leaky relu 或 elu (Exponential relu)，它们具有从这个问题中恢复的固有机制

随着这些变化，模型的损失降至

【讨论】：