GrammarCNN

A Grammar-Based Structural CNN Decoder for Code Generation

Our paper: https://arxiv.org/pdf/1811.06837.pdf

Usage

To train a new model

In folder model/, to train a new model

python3 run.py train [dev|test] [tree|var|func]

tree for nonterminal nodes, var for variable nodes, and func for function nodes.

dev and test denote the evaluation set.

To predict

After tree, var, and func were trained.

In folder predict/

python3 run.py [pre|eval]

Dependenices

• NLTK 3.2.1
• Tensorflow 1.3.1
• Python 3.5
• Ubuntu 16.04
• Java 1.8

Examples

We successly generated.

class BootyBayBodyguard(MinionCard ) : 
   def __init__ (self) :
       super().__init__("Booty Bay Bodyguard", 5, CHARACTER_CLASS.ALL, CARD_RARITY.COMMON)
   def create_minion (self, player) :
       return Minion(5, 4, taunt = True)

Example Code:

class BootyBayBodyguard(MinionCard ) : 
    def __init__ (self) :
        super().__init__("Booty Bay Bodyguard", 5, CHARACTER_CLASS.ALL, CARD_RARITY.COMMON)
    def create_minion (self, player) :
        return Minion(5, 4, taunt = True)

Our model tends to generate a structrual correct code, which leads to a higher StrAcc but a similar BLEU compared with previous works.

Code we generated.

class AnnoyoTron(MinionCard ) : 
    def __init__ (self) :
        super().__init__("Annoy-o-Tron", 2, CHARACTER_CLASS.ALL, CARD_RARITY.COMMON, minion_type = MINION_TYPE.MECH, divine_shield = True)
    def create_minion (self, player) :
        return Minion(1, 2, taunt = True, divine_shield = True)

Example Code:

class AnnoyoTron(MinionCard ) : 
    def __init__ (self) :
        super().__init__("Annoy-o-Tron", 2, CHARACTER_CLASS.ALL, CARD_RARITY.COMMON, minion_type = MINION_TYPE.MECH)
    def create_minion (self, player) :
        return Minion(1, 2, divine_shield = True, taunt = True)