CAM-loss Implementation by PyTorch

CAM loss unofficial implementation by pytorch

Run

# train
python run.py

python run.py --use_cam=True

# test
python run.py --mode=test

python run.py --use_cam=True --mode=test

Result

I think CAM loss is not effective in simple image or simple model.
There are also limitations in the model architecture for applying cam loss.
To know experimental results more accurately, maybe I need more experiments is required by changing hyper parameter(eg. alpha, model architecture, ...)

without cam loss

best acc : 0.9880
best f1 : 0.9883

with cam loss

best acc : 0.9873
best f1 : 0.9877