jungyhuk / latent-attention

Latent Attention For If-Then Program Synthesis

This repo provides the code to replicate the experiments in the paper

Xinyun Chen, Chang Liu, Richard Shin, Dawn Song, Mingcheng Chen, Latent Attention For If-Then Program Synthesis , in Proc. of NIPS 2016

Paper [arXiv] [NIPS]

Prerequisites

Tensorflow version >= v0.7

sklearn

jsonnet

Datasets

IFTTT

We use the same crawler from Quirk et al. to crawl recipes from IFTTT.com.

Processed data can be found in here.

Zapier

We additional provide a preprocessed dataset derived from Zapier recipes crawled using a crawler.

Processed data can be found under this folder.

Usage

Model architectures

The code includes the implementation of following models:

• BDLSTM+LA: in configs/model.jsonnet, set model/name to be "rnn", model/decoder to be "LA".
• BDLSTM+A: in configs/model.jsonnet, set model/name to be "rnn", model/decoder to be "attention".
• BDLSTM: in configs/model.jsonnet, set model/name to be "rnn", don't set model/decoder(delete this line or set it to "").
• Dict+LA: in configs/model.jsonnet, set model/name to be "Dict", model/decoder to be "LA".
• Dict+A: in configs/model.jsonnet, set model/name to be "Dict", model/decoder to be "attention".
• Dict: in configs/model.jsonnet, set model/name to be "Dict", don't set model/decoder(delete this line or set it to "").

Run experiments

In the following we list some important arguments in train.py:

• --dataset: path to the preprocessed dataset.
• --load-model: path to the pretrained model (optional).
• --config: path to the file that stores the configuration of model architecture.
• --logdir: path to the directory that stores the models (optional).
• --output: name of the file that stores the prediction results (no need to specify the filename extension, the output is a pickle (.pkl) file).

python train.py --dataset dataset/IFTTT/msr_data.pkl --config configs/model.jsonnet --logdir model --output result

To ensemble results of several models:

python test_ensemble_probs.py --data dataset/IFTTT/msr_data.pkl --res result_0.pkl result_1.pkl ... result_N.pkl

Citation

If you use the code in this repo, please cite the following paper:

@inproceedings{chen2016latent,
  title={Latent Attention For If-Then Program Synthesis},
  author={Chen, Xinyun and Liu, Chang and Shin, Richard and Song, Dawn and Chen, Mingcheng},
  booktitle={Proceedings of the 29th Advances in Neural Information Processing Systems},
  year={2016}
}

latent-attention's Issues

LabelEncoder()

Where is the LabelEncoder() that can be used to map prediction numbers to real instructions.
E.G.
'label_names': ['Instagram',
'Instagram.Any_new_photo_by_you',
'Dropbox',
'Dropbox.Add_file_from_URL'],
'labels': [45, 193, 11, 20],

Where can I find the map between label_names and labels ?

Thanks a lot. I am trying to use your code to play with prediction.

Format of Pre-processed data

Hello.
What is the format of the pre-processed dataset of IFTTT ?
On unpickling the msr_data.pkl file, the data in it is not making much sense to me... because the attributes like correct_action_param, correct_trigger_param, label_names, words, url etc. are occurring together again and again, as a bundle, and after them is another bundle of some other attributes , namely label_types, trigger_chans, action_chans, action_funcs etc and then is a bundle of action and trigger fields, which is followed by some fields which I don't understand but I think are the fields which require user input and are initialized as NULL and then... I don't understand.

Can you tell me how to make sense of the data ? Also, it would be pretty helpful if you can tell me how to see the data in a pretty-formatted format.