This experiment is a practice of neural network based on my current knowledge, in which I utilized the CNN network to carry out a classification task on cifar10 database.
In this experiment, I built an 8-layer network, which consists of 5 convolutional layers and 3 linear layers. The method Batch normalization was adopted.
visual explanation(image from internet)
architecture of the network (without resnet):
I also tested the resnet architecture, which worked pretty well in the task and gained a 68% accuracy.
my_batchnorm.py is an attempt to construct a functioning batchnorm module and probe into the mechanism of batch normalization.