nlil's Introduction

Neural Logic Inductive Learning

This is the implementation of the Neural Logic Inductive Learning model (NLIL) proposed in the ICLR 2020 paper: Learn to Explain Efficiently via Neural Logic Inductive Learning. The Transformer implementation is based on this repo.

Requirements

python 3.6+
pytorch 1.1.0+
numpy
tqdm

Knowledge completion on WN18 and FB15K

You can run knowledge completion task on WN18 and FB15K with provided scripts

bash run_wn.sh
bash run_fb.sh

Object classification on Visual Genome

First, download the scene-graph dataset from the official site (click "Download Scene Graphs")

https://cs.stanford.edu/people/dorarad/gqa/download.html

Extract the files, and run the following script to generate the dataset

bash preprocess.sh path/to/the/sgraph/folder

Now you can run object classification with

bash run_gqa.sh

Reference

@inproceedings{
    yang2020learn,
    title={Learn to Explain Efficiently via Neural Logic Inductive Learning},
    author={Yuan Yang and Le Song},
    booktitle={International Conference on Learning Representations},
    year={2020},
    url={https://openreview.net/forum?id=SJlh8CEYDB}
}

nlil's People

Contributors

Stargazers

Watchers

nlil's Issues

Questions about model implementation

Hi, I am very interested in your work.But I have some questions

The Operator search of Section 4 in the Paper is as follows

the EncoderLayer(nn.Module) take Q and V as the input .

However，the EncoderLayer(nn.Module) in your code dose not use Q as the input. The code just take the V as the input .

In addition, you use Q instead of q as input in DecoderLayer(nn.Module).

Is there a mistake in my understanding? Can you answer this doubt for me?

Issue running evensucc10 dataset

By simply changing the learning folder to the evensucc10 dataset through:

--data_root ../data/evensucc10

yields an error:

  File "<...>/model/SubLayers.py", line 45, in forward
    q = q.permute(2, 0, 1, 3).contiguous().view(-1, len_q, d_k)  # (n*b) x lq x dk
RuntimeError: cannot reshape tensor of 0 elements into shape [-1, 0, 32] because the unspecified dimension size -1 can be any value and is ambiguous

I believe it is caused by

mldel/Models.py: line 200~204: # TODO: debug ...

Can you help me take a look into this issue?

inference

Can you please provide a simple example of how to work with the trained model? As I understand it, once trained on a KB, it should generate FOL rules? Does it save them somewhere? Or we can get them interactively from the model?

Recommend Projects