Common Deep Learning Framework hides the backpropagation part of the neural network
in the frontend, this is very convenient but sometimes makes it hard to understand,
this repo implements the Convolution Neural Network and BatchNorm layer's backpropagation from scratch, using pytorch just for gpu acceleration and didn't use the automatic backprop function of the library.
It's training on the CIFAR10 dataset
after training 3 epochs(about 5 minutes on a RTX3060GPU),
you should get about 67 % test accuracy
python _1_train.py
python _2_test.py