what‘s meaning about 15 in cifar100

sahagobinda / GPM

Official [ICLR] Code Repository for "Gradient Projection Memory for Continual Learning"

MIT License

86 stars 16 forks source link

what‘s meaning about 15 in cifar100 #8

Closed muyuuuu closed 2 years ago

muyuuuu commented 2 years ago

what's meaning about 15, and why fill zero to parameters' gradient?

https://github.com/sahagobinda/GPM/blob/1a238ec9d2ca30bae8fd1707f161cc6bd093c72a/main_cifar100.py#L150-L157

sahagobinda commented 2 years ago

K<15 is used to exclude gradient projections of the final classifier layer. fill_(0) is used to stop gradient updates of the batch-norm parameters in the model after task 1.

muyuuuu commented 2 years ago

thank you~