An unofficial PyTorch implementation of the 2021 paper Pointer Value Retrieval: A new benchmark for understanding the limits of neural network generalization.
Both the datasets and the models are pure pytorch modules.
For training we used Pytorch-Lightning.
In order to train a model run the following command in the terminal
python train.py -c <path_to_config_file> -s <seed value>
Example configs can be found in the configs folder.
To recreate figures from the paper run the appropriate bash script:
massive_datasets_fig12.sh
massive_datasets_fig13_m1.sh
massive_datasets_fig13_m2.sh
The code requires integration with WandB, but it can be easily edited out.