Install requirements and tweak code

1. conda env create -f env.yml <- this is the environment the code was run in
2. In the conda environment, search for the captum package (e.g. under <env_name>/lib/python3.9/site-packages/), and in captum/_utils/gradient.py, modify the line grads = torch.autograd.grad(torch.unbind(outputs), inputs) to grads = torch.autograd.grad(torch.unbind(outputs), inputs, retain_graph=True, create_graph=True). This will create the computation graph for gradients and gradient-based attributions.
3. In the conda environment, search for the transformers package (e.g. under <env_name>/lib/python3.9/site-packages/), and in transformers/models/bert/modeling_bert.py, modify the line pooled_output = self.activation(pooled_output) to pooled_output = torch.tanh(pooled_output). This is in the BertPooler class, around line 660.
4. In the conda environment, search for the transformers package (e.g. under <env_name>/lib/python3.9/site-packages/), and in transformers/models/roberta/modeling_roberta.py, modify the line pooled_output = self.activation(pooled_output) to pooled_output = torch.tanh(pooled_output). This is in the RobertaPooler class, around line 578.

Run training

1. Set exp_folder parameter in scripts/[van,adv,far]_config.json as the desired logging folder
2. Set dataset, model, candidate_extractor... parameters in the config files
3. Run scripts/[van,adv,far]_train.py

Evaluate model

1. Set model_path parameter in scripts/[van,adv,far]_config.json to the trained model path
2. Set only_eval parameter in the same configuration file
3. Run scripts/[van,adv,far]_train.py, this will run the training script in evaluation mode