Question

这是我在Keras上使用ResNet34对Mnist进行的两个训练设置

两个网络之间的唯一区别在于“重新缩放”

这是我的两个问题：

为什么第一个网络通过两个重新缩放过程具有更好的性能？我的意思是像“ x_train / 255”和“ rescale = 1./255”

2。为什么我应该在开始训练时使用“ x_train * 255”。（下面的代码）

非常感谢您！

数据集：Mnist数据集环境：Colab 框架：Keras 型号：ResNet34

我只告诉大家两个网络之间的不同部分下面代码的训练结果：

x_train = x_train.reshape(-1, 28, 28, 1)
x_test = x_test.reshape(-1, 28, 28, 1)
x_train = x_train / 255 
x_test = x_test /255 

from keras.preprocessing.image import ImageDataGenerator
img_gen = ImageDataGenerator(
    rescale=1./255,
    horizontal_flip=False
    )

Epoch 1/5
3750/3750 [==============================] - 243s 65ms/step - loss: 0.2988 - acc: 0.9291 - val_loss: 0.0800 - val_acc: 0.9799
Epoch 2/5
3750/3750 [==============================] - 232s 62ms/step - loss: 0.1247 - acc: 0.9686 - val_loss: 0.0988 - val_acc: 0.9725
Epoch 3/5
3750/3750 [==============================] - 231s 62ms/step - loss: 0.0867 - acc: 0.9776 - val_loss: 0.0677 - val_acc: 0.9794
Epoch 4/5
3750/3750 [==============================] - 232s 62ms/step - loss: 0.0695 - acc: 0.9821 - val_loss: 0.0302 - val_acc: 0.9914
Epoch 5/5
3750/3750 [==============================] - 256s 68ms/step - loss: 0.0493 - acc: 0.9868 - val_loss: 0.0369 - val_acc: 0.9882

这是第二个网络结果是

x_train = x_train.reshape(-1, 28, 28, 1)
x_test = x_test.reshape(-1, 28, 28, 1)


from keras.preprocessing.image import ImageDataGenerator
img_gen = ImageDataGenerator(
    rescale=1./255,
    horizontal_flip=False
    )
Epoch 1/5
3750/3750 [==============================] - 264s 70ms/step - loss: 0.4926 - acc: 0.8603 - val_loss: 11.0879 - val_acc: 0.3114
Epoch 2/5
3750/3750 [==============================] - 230s 61ms/step - loss: 0.1993 - acc: 0.9448 - val_loss: 12.4943 - val_acc: 0.2240
Epoch 3/5
3750/3750 [==============================] - 230s 61ms/step - loss: 0.1596 - acc: 0.9570 - val_loss: 14.5385 - val_acc: 0.0980
Epoch 4/5
3750/3750 [==============================] - 231s 62ms/step - loss: 0.1090 - acc: 0.9694 - val_loss: 14.4918 - val_acc: 0.1009
Epoch 5/5
3750/3750 [==============================] - 229s 61ms/step - loss: 0.0848 - acc: 0.9769 - val_loss: 14.5740 - val_acc: 0.0958

我的第二个问题是为什么我应该添加“ x_train * 255”，因此“ * 255”在最后一部分中如何工作？

History = Resnet34_model.fit_generator(img_gen.flow(x_train*255, y_train, batch_size = 16),
                                      steps_per_epoch = len(x_train)/16, validation_data = (x_test,y_test), epochs = 5 )

如何理解在Keras上两次缩放至1/255以获得更好的结果

0 个答案: