I tried to implement CNN from scratch only using numpy.
Unfortunately, I couldn't solve the problem of "exploding gradients".
However, I want to show that the layers are working. I constructed smaller CNN. I also put actual network structure for this problem. It gives error. I have a detailed explanation for this subject in the report. However, if you want to see the error come from actual network just comment first network list and uncomment second network in the main.py file.
Also there is a code for sanity check written in Keras.