/blackout

Primary LanguagePython

blackout

torch 1.4.0 python 3.5

baseline:

python main_cot.py --gpu 4 --sess softmax.1 --cifar 100 -e 200 --lr 0.1 (test acc: 74.11%)

blackout:

python main_cot.py --gpu 4 --sess black1.50.1 --cifar 100 -e 200 --lr 0.1 --k 50 --blackout 1 (test acc: 75.42%)

python main_cot.py --gpu 4 --sess black3.50.1.1 --cifar 100 -e 200 --lr 0.1 --k 50 --blackout 3 --p 0.1 (test acc: 75.28%)

python main_cot.py --gpu 4 --sess resnextblack1.50 --cifar 100 -e 200 --lr 0.1 --k 50 --blackout 1 (test acc: 78.54%)

python main_cot.py --gpu 4 --sess resnextblack3.25.1 --cifar 100 -e 200 --lr 0.1 --k 25 --blackout 3 --p 0.1 (test acc: 79.1%)

--gpu wich gpu to use

--sess name to be saved as

--cifar (0,10,100) specify which dataset( mnist, cifar10,cifar100)

--blackout (0 , 1, 2, 3) enable different blackout versions (1 best so far, 3 is nonuniform sampling)

--GCE enbale COT training

--p probability to generate matrix