Dropblock Versions Save

Implementation of DropBlock: A regularization method for convolutional networks in PyTorch.

0.3.0

5 years ago

Changes:

  • Performance improvements of block mask computation (#21)
  • Performance improvements of bernoulli mask computation (#22)

Bug Fixes:

  • Fixed bug where too big block sizes lead to wrong block mask size (#20)

Notes:

  • Should be now compatible with PyTorch v1.0

0.2.1

5 years ago

Fixes a minor issue with the DropBlock2D block mask generation.

0.2.0

5 years ago

Breaking changes:

  • DropBlock is now DropBlock2D
  • feat_size argument no longer exists, now dinamically computed

New Features:

  • DropBlock3D
  • LinearScheduler
  • ResNet9 on CIFAR10 example script

Bug Fixes:

  • Refactored forward pass of DropBlock
  • FIxed gamma computation
  • Fixed CUDA compatibility

0.1.0

5 years ago

Index based DropBlock implementation