Due to different versions of torch, you may need to replace "cls_loss = args.cls_lambda * F.cross_entropy(forward_logit, labels)" with "cls_loss = args.cls_lambda * F.cross_entropy(forward_logit, labels.long())"
This repo contains Pytorch implementation of MCD (NeurIPS 2023 paper: D-Separation for Causal Self-Explanation). Most of our code are built on top of our previous work FR.
If the code has any bugs, please open an issue. We will be grateful for your help.
torch 1.13.1+cu11.6.
python 3.7.16.
RTX3090
Please refer to FR.
For the appearance aspect with sparsity being about 20%, run:
python -u decouple_bcr.py --correlated 1 --data_type beer --lr 0.0001 --batch_size 128 --gpu 0 --sparsity_percentage 0.175 --epochs 150 --aspect 0
If you have any other questions, please send me an email. I am happy to provide further help if you star this repo.
Preparing the code is really tedious, I will be appreciate if you star this repo before cloning it.
Please refer to FR.
The code is largely based on Car and DMR. Most of the hyperparameters (e.g. the '--cls_lambda'=0.9) are also from them. We are grateful for their open source code.