Run CutMix

[adidigit]

see issue #1

[adidigit]

I started configuring the parameters like the paper suggests with but with resnet18 like evgeni said in the paper they used pyramidNet

[adidigit]

Best accuracy (top-1 and 5 error): 36.43 11.59

[adidigit]

this run was with beta=0 so it didn't so cut mix, i will move this results to the plain resnet run

to run cut mix with resnet 18 like the paper:

python train.py \ --net_type resnet \ --dataset cifar100 \ --depth 18 \ --alpha 240 \ --batch_size 64 \ --lr 0.25 \ --expname CutMix \ --epochs 300 \ --beta 1.0 \ --cutmix_prob 0.5 \ --no-verbose

[adidigit]

Best accuracy (top-1 and 5 error): 36.68 12.24

[adidigit]

python train.py \ --net_type resnet \ --dataset cifar100 \ --depth 110 \ --alpha 240 \ --batch_size 64 \ --lr 0.25 \ --expname resnet110 \ --epochs 300 \ --beta 1.0 \ --cutmix_prob 0.5 \ --no-verbose

[adidigit]

Best accuracy (top-1 and 5 error): 22.13 5.3 ( in the paper : 20.11 4.43)

[adidigit]

tmux 1

python train.py \ --net_type resnet \ --dataset cifar100 \ --depth 101 \ --alpha 240 \ --batch_size 64 \ --lr 0.25 \ --expname resnet101CutMix \ --epochs 300 \ --beta 1.0 \ --cutmix_prob 0.5 \ --no-verbose

[adidigit]

26.71 6.7

[adidigit]

@naama-alon link to the checkpoints (last and best)

https://drive.google.com/drive/folders/1ejHIMiscwo1s2YSnEVnUOF1TjERlpEP5?usp=sharing