The results are available at https://wandb.ai/yuqinzhou/ATNLP?workspace=user-yuqinzhou. Here are the name conventions: "simple_split_hyper" is the hyperparameter search for exp1 using the transformer
"len_split_hyper" is the hyperparameter search for exp2 using the transformer
For the name like this "Exp1_var02_overall_0", it means the first run using the overall best models trained based on the 2% proportion of the Exp1 datasets.
To reproduce the results, the "transformer" folder contains:
Name | Description |
---|---|
ATNLP_transformer_hyper | Hyperparameter search for exp1 and exp2 using the transformer |
ATNLP_simple_32p | Exmaple code of the overall best model trained based on the 32% proportion of the Exp1 |
transformer_pos | transformer models with fixed positional embeddings |
transformer_new | transformer models with learnable positional embeddings |
The "Reimplementation" folder contains:
Name | Description |
---|---|
Exp1 | Simple split and random subset models |
Exp2 | Length split |
Exp3a | Models using "turn left" primitive |
Exp3b | Models using "jump" primitive |
Exp3c | Error inspection of experiment 3a and 3b |
Exp3d | Cosine similarity |
Exp3e | Model with different amount of compositional jump commands |