/BIC

pytorch implementation of Large Scale Incremental Learning

Primary LanguagePython

BIC

An unofficial pytorch implementation of "Large Scale Incremental Learning" from https://arxiv.org/abs/1905.13260

Dataset

Download Cifar100 dataset from https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz

Put meta, train, test into ./cifar100

Train

python main.py

Result

20 40 60 80 100
Paper 85.20 74.59 66.76 60.14 55.55
Implementation 83.80 68.75 63.50 58.25 54.93

Alpha & Beta

Adam (Bias correction layer)

20 40 60 80 100
Alpha 1.0 0.788 0.718 0.700 0.696
Beta 0.0 -0.289 -0.310 -0.325 -0.327

SGD (Bias correction layer)

20 40 60 80 100
Alpha 1.0 1.006 1.017 0.976 0.983
Beta 0.0 -2.809 -3.496 -3.447 -3.683

Different Optimizers make difference in alpha and beta.