CNN with tensorflow with MNIST
input of train_CNN.py is a pickle file pkl contain six data :trainX, validationX, testX, trainY, validationY, testY (only use train and validation) all of those are list,and i translate them to np in train_CNN.py i save the model in /workspace/checkpoint/CNN/ whose accuracy of validation is best
input of test_CNN.py is alse a pickle file pkl contain six data :trainX, validationX, testX, trainY, validationY, testY (only use test) and the output of of test_CNN.py is the average loss and accuracy of testX
CNN model with 2 conv layer and 2 max_pool, 1 fully connect layer and a classifier layer both conv use a kernel with size 5,5 and stride is 1 conv1 has 32 filters, conv2 has 64 filters both pooling use a kernel with 2,2 and stride is 2 before fc_layer, i reshape the output of pool2 layer fully connect layer return a vector with 512 dimension
input conv1_output pooling1_output conv2_output pooling2_output reshape fc_output num_class
1@28 * 28 -> 32@24 * 24 -> 32@12 * 12 -> 64@8 * 8 -> 64@4 * 4 -> 64 * 4 * 4 -> 512 -> 10
after 20 epoch , the accuracy on test is 99.2%