Learning to sort numbers using a seq2seq model.
Call pip install -r requirements.txt to install all dependencies.
All data can be generated using generate.py
python generate.py \
--name="train" \
--size=10000 \
--max_val=256 \
--min_length=2 \
--max_length=256 \
Models can be trained by setting appropriate parameters in experiments/train.py and then setting train.run() to be called within main.py, and last calling python main.py (yes that is quite a mouthful, I apologize for not configuring command line arguments).
Training for 1 epoch on the dataset generated from the example call above took about 10 minutes.
After training the model on train.txt, use generate.py to generate a test set (with name="test"), then run experiments/evaluate.py in the same way as was described for experiments/train.py above to see some example evaluation of the model.
The same as for evaluation, run experiments/reproduce.py via python main.py, and you should see similar results (for pointer).
Permutation: 0.0
Nondecreasing: 0.093
(for vanilla attention decoder)
Permutation: 0.0
Nondecreasing: 0.653
(note: I did very minimal fine-tuning, and a very short training scheme, so it is quite likely performance could exceed this)
Change RANDOM_SEED in utils.py to try a different shuffle (one could also generate a new dataset and train again.)
Please refer to Issues for currently known issues and areas of improvement for this code base.
I would like to acknowledge parts of this codebase (as indicated at the top of file) are modified from the following sources:
Last, the following implementation of Pointer Networks was used as a guiding reference for my own implementation.