IBM/cpg

code for the compositional program generator (cpg)

For more details see the paper: Compositional Program Generation for Few-Shot Systematic Generalization

https://arxiv.org/abs/2309.16467

Installation

pip install -r requirements.txt

Training

To train on SCAN

python -m src.model.train --save-dir pretrained/scan --dataset SCAN --training-set scan_data/SCAN_add_jump_0_train_no_jump_oversampling_extreme_few_shot.txt --validation-set scan_data/SCAN_add_jump_4_test.txt

To train on COGS

python -m src.model.train --save-dir pretrained/cogs --dataset COGS --training-set cogs_data/cogs_train_extreme_few_shot.tsv --validation-set cogs_data/cogs_dev.tsv

Optional arguments

--seed 1 to set the random seed to 1

--verbose to get detailed output and evaluation

--wandb to use wandb

Evaluation

To evaluate on SCAN

python -m src.model.evaluate --model-path pretrained/scan/<your-saved-model>.pkl --save-dir pretrained/scan --dataset SCAN --training-set scan_data/SCAN_add_jump_0_train_no_jump_oversampling_extreme_few_shot.txt --test-set scan_data/SCAN_add_jump_4_test.txt

To evaluate on COGS

python -m src.model.evaluate --model-path pretrained/cogs/<your-saved-model>.pkl --save-dir pretrained/cogs --dataset COGS --training-set cogs_data/cogs_train_extreme_few_shot.tsv --test-set cogs_data/cogs_gen.tsv