Counterfactual Explanation Based on Gradual Construction for Deep Networks

⭐ Star us on GitHub — it helps!!

PyTorch implementation for Counterfactual Explanation Based on Gradual Construction for Deep Networks

Install

You will need a machine with a GPU and CUDA installed.
Then, you prepare runtime environment:

pip install -r requirements.txt

If you get an error message ModuleNotFoundError: No module named 'urllib3' on conda enviroment, please try to install spacy like this:

conda install spacy==2.3.2

Use

MNIST dataset

Run the collowing command to get a counterfactual explanation for MNIST data.

python main.py --dataset=mnist --model_path=./models/saved/mnist_cnn.pt --data_path=example/MNIST/0.png --d=4 --target_prob=0.9

Results are saved in result folder.

IMDB dataset

For IMDB dataset, you should download 'en' model. Type following command.

python -m spacy download en

Run the collowing command to get a counterfactual explanation for IMDB data.

python main.py --dataset=imdb --model_path=./models/saved/tut4-model.pt --data_path="This film is good" --d=1 --target_prob=0.9 --n_iter=700

HELOC dataset

Run the collowing command to get a counterfactual explanation for HELOC data.

python main.py --dataset=heloc --model_path=./models/saved/MLP_pytorch_HELOC_allRemoved.pt --data_path=./example/HELOC/1.csv --d=1 --target_prob=0.7

The target probability over 0.7 is not allowed because of the pre-trained model capacity.

Arguments:

• dataset - Choose the experiment dataset
  • Available list: ['mnist','imdb','heloc']
• data_path - Input data (path)
• l2_coeff - Coefficient of the l2 regularization
• tv_beta - Exponential number of total variation (TV) regularization
• tv_coeff - Coefficient of the TV regularization
• n_iter - Iteration number
• lr - Learning rate
• target_class - Choose the target class
  • 0: a class that has the first highest proability from an original data
  • 1: a class that has the second highest proability from an original data
  • 2: a class that has the third highest proability from an original data
• target_prob - Target probability of the target class
• d - Determine size of mask
• model_path - Saved model path
  • Available list=['mnist_cnn.pt',tut4-model.pt','MLP_pytorch_HELOC_allRemoved.pt']

Understanding my paper

✅ Check my blog!! HERE