本文将和大家一起一步步尝试对Fashion MNIST数据集进行调参,看看每一步对模型精度的影响。（调参过程中，基础模型架构大致保持不变）
废话不多说，先上任务：
模型的主体框架如下（此为拿到的原始代码,使用的框架是keras）：
裸跑的精度为
--
将epochs由1变为12
当加上dropout之后时，model的结构如下：
=================================================================
layer_conv_1 (Conv2D) (None, 28, 28, 32) 320
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 14, 14, 32) 0
_________________________________________________________________
layer_conv_2 (Conv2D) (None, 14, 14, 64) 18496
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 7, 7, 64) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 7, 7, 64) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 3136) 0
_________________________________________________________________
dense_1 (Dense) (None, 64) 200768
_________________________________________________________________
dropout_2 (Dropout) (None, 64) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 650
=================================================================
Total params: 220,234
Trainable params: 220,234
Non-trainable params: 0
此时的精度为：
48384/55000 [=========================>....] - ETA: 0s - loss: 0.2562 - acc: 0.9068
49536/55000 [==========================>...] - ETA: 0s - loss: 0.2573 - acc: 0.9065
50688/55000 [==========================>...] - ETA: 0s - loss: 0.2572 - acc: 0.9064
51840/55000 [===========================>..] - ETA: 0s - loss: 0.2570 - acc: 0.9065
52992/55000 [===========================>..] - ETA: 0s - loss: 0.2574 - acc: 0.9063
54144/55000 [============================>.] - ETA: 0s - loss: 0.2573 - acc: 0.9062
55000/55000 [==============================] - 3s 50us/step - loss: 0.2574 - acc: 0.9063 - val_loss: 0.2196 - val_acc: 0.9192
Test loss: 0.2387163639187813
Test accuracy: 0.9133
Model saved
Select the first 9 images
加上dropout之后，对模型的分类精度还是有着不错的提升。提升了0.0276。
之后，尝试着将epoches增加，但对精度已经没有什么影响。
因此考虑数据增强，看看对分类精度的变化。代码如下：
#you can change the data_aug according to your need
datagen = ImageDataGenerator(
rotation_range=0.2,
zoom_range=0.2,
width_shift_range=0.2,
height_shift_range=0.2,
zca_epsilon=1e-6,
horizontal_flip=True,
)
model.fit_generator(datagen.flow(X_train, y_train, batch_size=batch_size),
steps_per_epoch=X_train.shape[0] // batch_size,
epochs=epochs,
validation_data=(X_test, y_test),workers=4,callbacks=[reduce_lr])
采取的epoches为24，因为epoches数太少，可能导致训练的不充分而使得精度不高，结果如下：
426/429 [============================>.] - ETA: 0s - loss: 0.5142 - acc: 0.8114
428/429 [============================>.] - ETA: 0s - loss: 0.5140 - acc: 0.8114
429/429 [==============================] - 22s 52ms/step - loss: 0.5140 - acc: 0.8115 - val_loss: 0.3480 - val_acc: 0.8751
Test loss: 0.3479545822381973
Test accuracy: 0.8751
Model saved
Select the first 9 images
可以看出，训练的精度为0.8115，验证的精度为0.8751。表明模型还有进步空间，离过拟合，还有一段距离。增加轮次和增加神经网络的节点数，对模型精度还能继续提高。
但是数据增强后，模型的收敛速度变慢了，精度也下降了些。
接下去，对模型学习率进行调整。初始学习率为1e-3
每12个epoches，学习率为原来的 1/10，总的epoches数为36，代码，结果如下：
def scheduler(epoch):
# 每隔12个epoch，学习率减小为原来的1/10
if epoch % 12 == 0 and epoch != 0:
lr = K.get_value(model.optimizer.lr)
K.set_value(model.optimizer.lr, lr * 0.1)
print("lr changed to {}".format(lr * 0.1))
return K.get_value(model.optimizer.lr)
reduce_lr = LearningRateScheduler(scheduler)
model.fit(X_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(X_val, y_val),callbacks=[reduce_lr])
53248/55000 [============================>.] - ETA: 0s - loss: 0.1977 - acc: 0.9287
54400/55000 [============================>.] - ETA: 0s - loss: 0.1981 - acc: 0.9287
55000/55000 [==============================] - 3s 53us/step - loss: 0.1980 - acc: 0.9288 - val_loss: 0.2026 - val_acc: 0.9266
Test loss: 0.22499265049099923
Test accuracy: 0.9216
Model saved
Select the first 9 images
精度还是有着不错的提升，达到了0.9216.
时间有限，文末将放所有完整的github代码链接: https://github.com/air-y/pythontrain_cnn
(PS:有兴趣的朋友可以自己调试下玩玩(基本保持2个卷积和全连接的结构不变,尽可能的提升模型精度)，我们交流下。真是好久没用keras了,一直在使用pytorch)