Exploiting Code Symmetries for Learning Program Semantics

See the ICML 2024 paper Exploiting Code Symmetries for Learning Program Semantics for more details.

Installation

pip install .

Training

Function Signature

bash command/finetune/finetune_arg_type_graph.sh $TRAIN_KEYWORD $VALID_KEYWORD $LABEL_TYPE $CUDA_VISIBLE_DEVICES

Memory Region

bash command/memory_region/finetune_symmetry_concrete.sh $CUDA_VISIBLE_DEVICES $MEMORY_REGION_CLASS $PERTURB

Function Similarity

bash command/similarity/pretrain_symmetry.sh $TRAIN_KEYWORD $VALID_KEYWORD $CUDA_VISIBLE_DEVICES

Function Name

bash command/source_code/pretrain_symmetry.sh $CUDA_VISIBLE_DEVICES

Testing

Function Signature

bash command/finetune/validate_arg_type_graph.sh $TRAIN_KEYWORD $VALID_KEYWORD $LABEL_TYPE $CUDA_VISIBLE_DEVICES

Memory Region

bash command/memory_region/eval_symmetry.sh $CUDA_VISIBLE_DEVICES $MEMORY_REGION_CLASS $PERTURB

Function Similarity

bash command/similarity/eval_symmetry.sh $TRAIN_KEYWORD $VALID_KEYWORD $CUDA_VISIBLE_DEVICES

Function Name

bash command/source_code/eval_symmetry.sh $CUDA_VISIBLE_DEVICES

Example Dataset

Memory Region: Google Drive

Function Similary: Google Drive

Function Name: Google Drive