The RobustFill network by Devlin et al. is trained for the following task -- based on a few example input-output string pairs, generate a program in a domain-specific language that transforms the given inputs into the given outputs. This program can then be used to transform unseen inputs.
This repo contains a PyTorch checkpoint of the model and optimizer state dicts after
4 gpus x 195,200 minibatches x 32 batch_size = 24,985,600 samples of training.
Please see the demo notebook for an example of how to use it:
Demo Notebook
