Neural Engineers TF

collaboration between: Ethan Caballero, Yad Faeq, Andy Zhang, Alex Skidanov

---

NE implementation in TensorFlow for learning to engineer.

This model is a decription2code task baseline (rnn-based seq2seq with AST decoder & adaptive number of attention hops per decode step) (with two stage SL then RL training) that yielded underwhelming results.

Acknowledgments

Barronalex's and Therne's tf-DMN+s,

Alexander Johansen's seq2seq modules,

Jon Gauthier's RML attempt & policy gradient module

MarkNeumann's adapative attention module

and seqGANs policy gradient modules.

Repository Contents

file	description
dmn_plus.py	contains the DMN+ model
dmn_train.py	trains the model on a specified task
dmn_test.py	tests the model on a specified task
get_data.sh	shell script to fetch & load Description2Code Dataset

---

Instructions to get data and run model: (Work In Progress)

1. The full problems dataset can be downloaded and pre-processed can be generated through:

  sh get_data.sh

To tweak for new directories:

2. The root directories can be set for the coding problems in yads_data_loader.py :

  rootdir1 = './description2code_current/codeforces_delete'
  rootdir2 = './description2code_current/hackerearth/problems_college'
  rootdir3 = './description2code_current/hackerearth/problems_normal'

3. Then tweak these parameters per needed experiment:

questions_count = 3000
answers_count = 50
max_len_words = 800
regular_desc = True

4. to run the model:

 python dmn_train.py

Note about the data/sample:

---

There's some toy data in tmp folder for if you don't feel like downloading full data. If you plan on using full data, delete all files in tmp folder before running sh get_data.sh and maybe before python dmn_train.py as well.

Results:

outputs always overfit to simplistic incorrect answers such as: a=raw_input(); print(2-a)

Interestingly, model learned dynamic adaptation of number attention hops (per each decode step) that seems to correlate with complexity of each decode step:

values in 1st line are number of attention hops for each decode step; values in 2nd line are output of each decode step. The two lines are aligned below for visualization purposes:

num of attn hops per dec step: array([ 1., 1., 6.,          1., 2., 1., 3.,       6., 2., 7.,  2.,  3.])
output of dec step:                    a   =   raw_input(       )   ;   print(    2   -   a     )   END

How project would be approached differently today:

-Transformer from "attention is all you need" https://arxiv.org/abs/1706.03762

-Use a Differentiable Programming Language as decribed in "Differentiable Functional Program Interpreters" https://arxiv.org/abs/1611.01988 so that you can scrap all score/value estimators from RL and just get the exact gradient of the reward with respect to the parameters

-Hindsight Experience Replay https://arxiv.org/abs/1707.01495

-unsupervisedly pretrain the encoder & decoder

-better/bigger data such as maybe the 150370 docstring2code pairs from https://github.com/EdinburghNLP/code-docstring-corpus or the 17000 programming challenge description2code pairs (original was just 7000) that Alex Skidanov recently acquired.

^if you're interested in trying any of this, message me or join near.ai