Try different kernel size like (2, 2), (3, 3), (5, 5) and stride size (2, 2), (1, 1)
I found out kernel size (3, 3) and stride size (1, 1) works best.
Most importance parameters:
Kernel Size and Stride Size
BN
Comparison | MLP | Ruochen's Network |
---|---|---|
Acc(5 epoch Train ACC) | 0.986465 | 0.998687 |
Acc(5 epoch CV) | 0.975500 | 0.992200 |
CPU 5 Epoch time | 2.144 | 361.695 |
GPU 5 Epoch time | NA | NA |
- https://bittigerimages.s3.amazonaws.com/gitbookImages/DS502-1702/DS502-1702%20%20Week3%20theory%20%28With%20Watermark%29%20%28Compressed%29-%E6%B0%B4%E5%8D%B0.pdf
- http://vda.univie.ac.at/Teaching/FDA/16w/A3/A3.pdf
- https://books.google.com/books?id=Vb0tDwAAQBAJ&pg=PA207&lpg=PA207&dq=mx+sym+convolution+function+example&source=bl&ots=Vd2MNarP6J&sig=DAM6fGYQW92hRgCdxVJbGCrcUjA&hl=en&sa=X&ved=0ahUKEwjxhqnz9J_VAhXHhVQKHT3CC_M4ChDoAQgtMAE#v=onepage&q=mx%20sym%20convolution%20function%20example&f=false