Karel Dataset

This is the code used to generate the Karel dataset as described in following papers:

• Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis (code)
• Neural Program Meta-Induction

Introduction

Karel is an educational programming languag. The Domain-specific language for Karel programs is:

but I added {, } and removed : to explicitly represent the scope without tabspace.

The state of the grid world are represented as a H x W x 16 tensor. Each cell of grid is the 16-dimensional vector indicated as:

Index	Description
0	Hero facing North
1	Hero facing South
2	Hero facing West
3	Hero facing East
4	Wall or not
5	0 marker
6	1 marker
7	2 marker
8	3 marker
9	4 marker
10	5 marker
11	6 marker
12	7 marker
13	8 marker
14	9 marker
15	10 marker

Usage

0. Installation

To install karel package:

$ pip install karel

1. Data generation

There are two types of parser which can be defined with --parser_type:

1. curly: Karel + curly braces ({, })
2. synthesis: Karel for program synthesis (examples can be found here)

To generate programs and its input/output examples:

$ python generate.py --data_dir=data --max_depth=5 --parser_type=synthesis

which will generate data/train.npz, data/test.npz and data/val.npz and you can use these data with:

data = np.load(npz_path)
for input, output, code in zip(data['inputs'], data[outputs'], data['codes']):
    train(input, output, code)

To generate programs only as text:

$ python generate.py --mode=text --beautify=True --parser_type=curly

and it will generate data/train.txt, data/test.txt and data/val.txt that contain codes like:

def run() {
  while(right_is_clear()) {
    repeat(7) {
      put_marker()
    }
  }
}
def run() {
  while(left_is_clear()) {
    repeat(13) {
      pick_marker()
    }
  }

  while(no_markers_present()) {
    ifelse(right_is_clear()) {
      put_marker()
    }
    else {
      pick_marker()
    }
  }
}

2. Interpreter

To run Karel interpreter (with random grid world):

$ python -m karel.parser_with_curly
In [1]: def run() {
   ...:   repeat(7) {
   ...:     ifelse(front_is_clear()) {
   ...:       move()
   ...:     }
   ...:     else {
   ...:       turn_right()
   ...:     }
   ...:   }
   ...: }
Input:  ########
        #.1.1#.#
        ##..1..#
        #.11.1##
        #1>1..1#
        ##....1#
        #.#....#
        ########
Output: ########
        #.1.1#.#
        ##..1..#
        #.11.1##
        #1.1..1#
        ##....1#
        #.#...v#
        ########
In [2]:

or,

$ python -m karel.parser_for_synthesis
In [1]: DEF run m( IF c( frontIsClear c) i( turnRight move i) m)
Input:  ########
        #....#.#
        #......#
        #.oo..##
        #o>....#
        #.....o#
        #......#
        ########
Output: ########
        #....#.#
        #......#
        #.oo..##
        #o.....#
        #.v...o#
        #......#
        ########
In [2]:

To run Karel interpreter with a world file (ex. assets/simple.world):

$ python -m karel.parser_with_curly --world=assets/simple.world
In [1]: def run() {
   ...:   repeat(7) {
   ...:     ifelse(front_is_clear()) {
   ...:       put_marker();
   ...:       move()
   ...:     }
   ...:     else {
   ...:       turn_right()
   ...:     }
   ...:   }
   ...: }
Input:  ########
        #....#.#
        ##..o..#
        #.....##
        #.>....#
        ##.....#
        #.#....#
        ########
Output: ########
        #....#.#
        ##..o..#
        #.....##
        #.ooooo#
        ##....o#
        #.#...v#
        ########
In [2]:

Author

Taehoon Kim / @carpedm20