Digit_recognizer

Using CNN, LeNet, AlexNet as backbone network and apply mixup to improve and compare.

Result

• Training set : Testing set = 4:1, separated from train.csv.
• MNIST dataset come from Kaggle competition.
• Alpha=0 means not using mixup
• For LeNet, changing alpha from 0 to 0.1, accuracy improved 0.9%!

Alpha	CNN	LeNet	Alexnet
0	98.631	97.631	98.821
0.1	99.024	98.524	98.440
0.2	98.964	98.262	98.774
0.4	99.060	98.464	99.000
0.8	98.857	98.476	98.952

Run

Just run the notebook, require Pytorch.

You can add alpha=0 to the mixup notebooks so you don't need to run the original ones.

I still put the original ones on for someone who don't want to use mixup.

• You may choose the GPU version CNN_GPU

Mixup

mixup: Beyond Empirical Risk Minimization