Class Activation Map(CAM) implementation by PyTorch. Paper can be seen in here.
Implementation detail is different from the original implementation from this repository
You can execute this program by type this in shell
cd cam && bash do.sh
for installing overall depency of this repo, follow this installation :
sudo apt update && sudo apt install build-essential
pip install chainercv && pip install matplotlib && pip install imageio && pip install pillow && pip install scikit-image
sudo apt-get remove cython && pip install -U cython && pip install git+https://github.com/lucasb-eyer/pydensecrf.git
for spend some lazy times... use tmux.
sudo apt-get install tmux
IRN & AffinityNet training is recommended to be done in single GPU. For regarding this Issue, MeanShift layer is dependent on batch size per GPU(similar as batchnorm in process), so the training environment is recommended to be set same as the original repository.
Last version of this repository has so many bugs, so it leads to borrow base codes from IRN implementation. Thanks for original repository writer. Applying denseCRF and AffinityNet(part of IRN) is heavily done from original repository.
use this command for treating the error on 'typing-extensions' :
pip install typing-extensions==4.3.0
use this command for treating the error on 'importing ToTensor in albumentation'
pip install albumentations==0.4.6