An implementation of convolution nueral network for recognizing hand-written digits using numpy.
Run main.py and wait for training process. Then, it will show the loss curve, accuracy curve, and some performance index. After closing the plot, you can enter an image index in the terminal to supervise the recognizing result directly.
(It uses MNIST built-in keras, but you can modify it easily.)
Fig.1 Loss curve and accuracy curve.
Fig.2 Learning rate curve
Fig.3 Preview
There is something wrong with weight updation in the process of convolution layer backward propagation. To prevent this problem, we use 2 Sobel filters and cancelled weight updation in the convolution layer. The result is acceptable as well. You can fix bugs inside ConvLayer on yourself.