Home

Awesome

This is a manual to reproduce the results of nsm (mapo) with a clean WikiTableQuestions download, which was not preprocessed by the authors of nsm. For the preprocessed version follow the instructions on NSM's (MAPO) repository on GitHub.

The manual also includes necessary adjustments for a custom dataset to run.

Prerequisites:

Preprocessing:

Training:

Edit 'run.sh' in neural-symbolic-machines/table/wtq to have the correct paths. Other information can be found on the project's GitHub page.
Notes:

  1. n_actors is heavy, on local I recommand changing it to 3.
  2. train_use_gpu/eval_use_gpu should be changed to 0 if you don't have CUDA or gpu tensorflow.
  3. Use train_gpu_id/eval_gpu_id for the id of the gpu to use (if you have multiple)
  4. Add debug flag to run if something went wrong.

Adjusting a custom dataset: