This is a PyTorch implementation of our ICLR 2021 paper:
Learning to Deceive Knowledge Graph Augmented Models via Targeted Perturbation
Mrigank Raman, Aaron Chan*, Siddhant Agarwal*, Peifeng Wang, Hansen Wang, Sungchul Kim, Ryan Rossi, Handong Zhao, Nedim Lipka, Xiang Ren.
ICLR 2021.
*=equal contritbution
Clone the repository
git clone https://github.com/INK-USC/deceive-KG-models.git
- Download the pretrained models and datasets
cd deceive-KG-models/obqa
bash scripts/download.sh
cd data/cpnet
wget https://csr.s3-us-west-1.amazonaws.com/tzw.ent.npy
- Preprocess the data
cd deceive-KG-models/obqa
python preprocess.py
By default, all available CPU cores will be used for multi-processing in order to speed up the process. Alternatively, you can use -p to specify the number of processes to use:
python preprocess.py -p 20
- Train the base classifier
Then train a grn model (for example)using the command:
python grn.py -ds obqa --encoder bert-base-uncased -bs 64 -mbs 4 -dlr 1e-3
Similarly train RN and GN as well.
- Train the triple classifier
python Get_neg_triples.py
python deep_triple_classifier.py
- Pruning the Graph for only the useful nodes
python new_graph.py
- Running Heuristics
python heuristics.py
The attributes are:
-np --num_pert: number of perturbations
--type type of perturbation(rel for Relation Swapping, edge for Edge Deletion and edge1 for Edge Rewiring)
For Relation Replacement use the command:
python train.py --mode_type eval --num_epochs 1 --save_dir ./saved_models/KG/model_25 --model_id 5 --enable_shuffle --dqn_lstm_len 100 --dqn_batch_size 16 --dqn_train_step 50 --log_path log_25.csv --steps_after_collecting_data 2000
- Training the RL agent
python train.py --mode_type train --num_epochs 1 --save_dir ./saved_models/KG/model_25 --model_id 1 --enable_shuffle --dqn_lstm_len 100 --dqn_batch_size 16 --dqn_train_step 50 --log_path log_25.csv --steps_after_collecting_data 2000
- Evaluating the RL agent
python train.py --mode_type eval --num_epochs 71801 --num_steps 70000 --save_dir ./saved_models/KG/model_25 --model_id 1 --debug_mode
what you have to change for specific model:
change of GPU number
change saved path to coincide with your saved model: ./saved_models/KG/model_25;
change num_steps: the number of steps you want to perturb, 70000 as default
cd deceive-KG-models/RippleNet
- Train the base classifier
bash scripts/run_movie.sh
or
bash scripts/run_music.sh
- Run Heurisitics
python main_RN.py
The attributes are:
--dataset: movie or music
--type : Type of perturbation
--num_pert : Number of perturbations
- Train RL agent
python train.py --mode_type train --num_epochs 1 --save_dir ./saved_models/KG/model_25 --model_id 1 --enable_shuffle --dqn_lstm_len 100 --dqn_batch_size 16 --dqn_train_step 50 --log_path log_25.csv --steps_after_collecting_data 2000
- Evaluate the RL agent
python train.py --mode_type eval --num_epochs 71801 --num_steps 70000 --save_dir ./saved_models/KG/model_25 --model_id 1 --debug_mode