Code for paper: Systematic Generalization on gSCAN with Language Conditioned Embedding

First clone https://github.com/LauraRuis/multimodal_seq2seq_gSCAN , enter multimodal_seq2seq_gSCAN/read_gscan and run

python read_gscan.py --dataset_path=../data/compositional_splits/dataset.txt --save_data --output_file=parsed_dataset.txt

Then move parsed_dataset.txt to parsed_dataset/ in this repo and run preprocess_parsed_dataset.py.

We use the same environment as the baseline, and dgl library.

python main_model.py

Model parameters are defined in model/config.py, some common settings can also be modified through command line.

• --run exp_name Set the experiment name
• --txt If enabled, the model redirects all of the outputs to exp/exp_name.txt
• --load path_to_model Load the checkpoint
• --baseline Switch the model to baseline

As part of our experiments, we ran two models on the same data split and compared their results. Results are saved in json format that is compatible with the visualization code from gSCAN.

Note: Remember to change the paths to models' checkpoints in model_compare.py before running. python model_compare.py

In eval_best_model.py.