davidsonic / interpretable_cnns_via_feedforward_design Goto Github PK

I run the following scripts
python Getkernel.py
python Getfeature.py
python Getweight.py
python mnist_test.py
But the testing acc is only 0.0974.

In MNIST/saab.py (154-157), you first remove the sample patches' feature means, then the patch means.
However, the order is reversed in CIFAR10/saab.py (109-113). Won't it affect the result?

Also, I am curious why in CIFAR10/saab.py (198 & 252), you perform the projection using the sample patches but not the patches that have removed the mean while you perform

transformed=np.matmul(sample_patches_centered, np.transpose(kernels))

in MNIST/saab.py.

First,thanks for your code. I have run the code
I don't find the feat.pkl , run Get_feature.py to get feat.pkl
when i run Get_kernel.py ,my server have has MemoryError (my server has 128g physical memory)
but , when i run about minst code don't happened memoryerror , It is happened at cifar code
So, I'd like to ask if you have any memoryerror in your experiment, I think 128g physical memory already big enough