import os os.environ["CUDA_DEVICE_ORDER"]&＃61;"PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"]&＃61;"1"import keras from keras.datasets import cifar10 from keras import backend as K from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Dense, BatchNormalization, Activation, MaxPooling2D from keras.models import Model from keras.layers import concatenate,Dropout,Flattenfrom keras import optimizers,regularizers from keras.preprocessing.image import ImageDataGenerator from keras.initializers import he_normal from keras.callbacks import LearningRateScheduler, TensorBoard, ModelCheckpointnum_classes &＃61;10 batch_size &＃61;64# 64 or 32 or other epochs &＃61;300 iterations &＃61;782 DROPOUT&＃61;0.5# keep 50% CONCAT_AXIS&＃61;3 weight_decay&＃61;1e-4 DATA_FORMAT&＃61;&＃39;channels_last&＃39;# Theano:&＃39;channels_first&＃39; Tensorflow:&＃39;channels_last&＃39; log_filepath &＃61;&＃39;./alexnet&＃39;
Total params: 21,622,154 Trainable params: 21,622,154 Non-trainable params: 0
5&＃xff09;开始训练
# set optimizer sgd &＃61; optimizers.SGD(lr&＃61;.1, momentum&＃61;0.9, nesterov&＃61;True) model.compile(loss&＃61;&＃39;categorical_crossentropy&＃39;, optimizer&＃61;sgd, metrics&＃61;[&＃39;accuracy&＃39;])# set callback tb_cb &＃61; TensorBoard(log_dir&＃61;log_filepath, histogram_freq&＃61;0) change_lr &＃61; LearningRateScheduler(scheduler) cbks &＃61;[change_lr,tb_cb]# set data augmentation datagen &＃61; ImageDataGenerator(horizontal_flip&＃61;True,width_shift_range&＃61;0.125,height_shift_range&＃61;0.125,fill_mode&＃61;&＃39;constant&＃39;,cval&＃61;0.) datagen.fit(x_train)# start training model.fit_generator(datagen.flow(x_train, y_train,batch_size&＃61;batch_size),steps_per_epoch&＃61;iterations,epochs&＃61;epochs,callbacks&＃61;cbks,validation_data&＃61;(x_test, y_test)) model.save(&＃39;alexnet.h5&＃39;)
6&＃xff09;结果分析 training accuracy 和 training loss test accuracy 和 test loss loss 和 accuracy 一个样&＃xff0c;正宗的耐克标志&＃xff01;超级过拟合&＃xff0c;fully connection 果然名不虚传
x &＃61; Conv2D(96,(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;,kernel_initializer&＃61;&＃39;uniform&＃39;)(img_input)# valid
其它部分代码同 alexnet
参数量如下&＃xff08;增加了挺多的&＃xff09;&＃xff1a;
Total params: 87,650,186 Trainable params: 87,650,186 Non-trainable params: 0
alexnet Total params: 21,622,154
因为 down sampling 的次数少了3次&＃xff0c;最后接 fc 层 的feature map 的 resolution 会大 8×8倍&＃xff0c;而且 CNN 的参数量大多都集中在 fc 层的结构&＃xff0c;所以参数量变多了可想而知&＃xff01; 叫 slim 的原因是缩减了一些 down sampling 层&＃xff0c;如果是全卷积网络&＃xff0c;参数量会减少或者不变&＃xff01;
test accuracy 和 test loss 美滋滋&＃xff0c;90%&＃43;了&＃xff0c;可是还是有一定的过拟合现象&＃xff01;
x &＃61; Conv2D(96,(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;,kernel_initializer&＃61;&＃39;he_normal&＃39;,kernel_regularizer&＃61;regularizers.l2(weight_decay))(img_input)# valid
其它代码同 alexnet_slim
参数量如下&＃xff08;不变&＃xff09;&＃xff1a;
Total params: 87,650,186 Trainable params: 87,650,186 Non-trainable params: 0
alexnet Total params: 21,622,154
alexnet_slim Total params: 87,650,186
结果分析如下 training accuracy 和 training loss test accuracy 和 test loss 精度上到了 92%&＃xff0c;过拟合现象得到了缓解
2.4 alexnet_thinner
alexnet_thinner_2 在 alexnet_slim_regular 的基础上&＃xff0c;将网络所有的 filters number 改为原来的 1/2&＃xff0c;包括 fully connection&＃xff0c; alexnet_thinner_4 在 alexnet_slim_regular 的基础上&＃xff0c;将网络所有的 filters number 改为原来的 1/4&＃xff0c;包括 fully connection。
其它代码同 alexnet_slim_regular
参数量如下&＃xff1a;
alexnet Total params: 21,622,154
alexnet_slim Total params: 87,650,186
alexnet_slim_regular Total params: 87,650,186
alexnet_thinner_2 Total params: 6,074,250
alexnet_thinner_4 Total params: 5,488,106
结果分析如下 train accuracy 和 train loss test accuracy 和 test loss 对比来看&＃xff0c;thinner_2 模型比较大&＃xff0c;有些过拟合了&＃xff0c;thinner_4 模型较小&＃xff0c;后面loss 没有上升的迹象&＃xff0c;直觉上的感觉是 convolution 和 fc 要相匹配&＃xff0c;convolution 过小&＃xff0c;fc 过大&＃xff0c;则会过拟合&＃xff0c;相匹配会好一些&＃xff08;可能 hyper parameters 不是最优&＃xff0c;所以结论并不靠谱&＃xff09;&＃xff01;
京公网安备 11010802041100号