This repo is a PyTorch cuda extension of the paper "CARAFE: Content-Aware ReAssembly of FEatures". It can replace the upsamling layer in the networks and is easy to be integrated into mmdetection.

CARAFE can be implemented in plain PyTorch code by torch.nn.functional.unfold. However, the unfold function reassembles the features within a local region to generate a large feature map, which causes memory problem and is slow when the input is large.

CARAFE cuda extension can solve the memory problem and is ~10x faster than the code in plain PyTorch. More details can be referred to test.py.

• Keep the environment to be consistent with mmdetection.
• Build

cd carafe_layer/
python setup.py build_ext --inplace

• Test

python test.py

We replace the upsampling layers with carafe_module in Mask R-CNN. The results with ResNet-50 are list as follows:

NOTE: M.H represents Mask Head in Mask R-CNN.

More results will be added.