HOUDINI - Lifelong Learning

Code for running the experiments, described in our NeurIPS 2018 paper "HOUDINI: Lifelong Learning as Program Synthesis" (https://arxiv.org/abs/1804.00218)

Requirements:

• Pytorch 0.3.1

Datasets

The datasets are automatically downloaded when needed.

Running

Houdini

For the counting sequences 1, 2, 3, and the long sequence:
usage: python HOUDINI/Eval/CS_LS.py [-h] [--synthesizer {enumerative,evolutionary}]
                             --taskseq {cs1,cs2,cs3,ls} [--dbg]

For graph the sequences:
usage: python HOUDINI/Eval/GraphSeq.py [-h] [--synthesizer {enumerative,evolutionary}]
                             --taskseq {gs1, gs2} [--dbg]

optional arguments:
  -h, --help            show this help message and exit
  --synthesizer {enumerative,evolutionary}
                        Synthesizer type. (default: enumerative)
  --taskseq {cs1,cs2,cs3,ls}
                        Task Sequence
  --dbg                 If set, the sequences run for a tiny amount of data

Baselines

Counting Sequence 1, Standalone: python Baselines/CountingSeqLongSeq.py cs1 sa

Counting Sequence 1, low-level-transfer: python Baselines/CountingSeqLongSeq.py cs1 wt

Counting Sequence 2, Standalone: python Baselines/CountingSeqLongSeq.py cs2 sa

Counting Sequence 2, low-level-transfer: python Baselines/CountingSeqLongSeq.py cs2 wt

Counting Sequence 3, Standalone: python Baselines/CountingSeqLongSeq.py cs3 sa

Counting Sequence 3, low-level-transfer: python Baselines/CountingSeqLongSeq.py cs3 wt

Long Sequence, Standalone: python Baselines/CountingSeqLongSeq.py ls sa

Long Sequence, low-level-transfer: python Baselines/CountingSeqLongSeq.py ls wt