My replication code for the paper Pointer Networks.
tldr: go to results
The official data is hosted at http://goo.gl/NDcOIG. It can be downloaded using gdown
$ gdown --folder --output data https://drive.google.com/drive/folders/0B2fg8yPGn2TCMzBtS0o4Q2RJaEUor manually using the web interface.
Alternatively, since gdown has sometimes failed me, I'm hosting it myself (hopefully it's not illegal). You can download it like so
$ mkdir -p data && wget -O- https://users.dcc.uchile.cl/~gchapero/datasets/ptr-nets-data.tar.gz | tar -C data -xzIn any case, the project root should have a data dir that looks like this:
data
├── convex_hull_10_test.txt
├── convex_hull_200_test.txt
├── ...
├── convex_hull_5_test.txt
├── README.txt
├── tsp_10_test_exact.txt
├── tsp_10_train_exact.txt
├── ...
└── tsp_5_train.zip
Not all data is available, only data for convex hull and tsp. Even then, some splits are missing. See here for available splits for convex hull and here for tsp splits.
Notice also that some splits in tsp were identified by computing the average solution distance, since not all of them are tagged with the algorithm used to solve the problem.
The code was tested on python 3.9 using pytorch 1.11. Other versions might work but your mileage may vary.
To install the dependencies run
$ pip install -r requirements.txtNote that the pytorch and cuda versions are pinned, so if you are using a gpu with different cuda capabilities you should modify the requirements file accordingly.
The cli has two commands, train and replicate
$ python -m ptrnets --help
Usage: python -m ptrnets [OPTIONS] COMMAND [ARGS]...
Options:
--help Show this message and exit.
Commands:
replicate
trainIf you want to replicate the results of the paper run
$ python -m ptrnets replicateThis should handle loading the data and running all the experiments of the paper
I also provide a simple train command where you can choose to either train the network
for convex hull or tsp, and tweak hyperparameters. Here is the synopsis for training in
convex hull.
$ python -m ptrnets train convex-hull --help
Usage: python -m ptrnets train convex-hull [OPTIONS]
Options:
--train-npoints [5|10|50|200|500|5-50]
[required]
--test-npoints [5|10|50|200|500|5-50]
[required]
--learn-rate FLOAT [default: 1.0]
--hidden-size INTEGER [default: 256]
--init-range FLOAT... [default: -0.08, 0.08]
--batch-size INTEGER [default: 128]
--max-grad-norm FLOAT [default: 2.0]
--seed INTEGER [default: 42]
--help Show this message and exit.These results were produced by running python -m ptrnets replicate --write, and
copy-pasted from the reports dir. The training curves can be seen in the experiment
at TensorBoard.dev.
The command took ~6h to run on a single RTX3090 (24gb). I didn't see more than ~5gb of VRAM usage.
Compare with Table 1 of the paper.
| method | trained n | n | accuracy | area |
|---|---|---|---|---|
| ptr-net | 50 | 50 | 66.1% | 100.0% |
| ptr-net | 5-50 | 5 | 93.8% | 99.8% |
| ptr-net | 5-50 | 10 | 87.9% | 99.9% |
| ptr-net | 5-50 | 50 | 63.1% | 100.0% |
| ptr-net | 5-50 | 200 | 10.1% | 99.9% |
| ptr-net | 5-50 | 500 | 0.1% | 99.6% |
Compare with Table 2 of the paper.
| n | ptr-net |
|---|---|
| 5 | 2.12 |
| 10 | 2.88 |
| 50 (a1 trained) | 6.7 |
| 5 (5-20 trained) | 2.18 |
| 10 (5-20 trained) | 3.05 |
| 20 (5-20 trained) | 4.23 |
| 40 (5-20 trained) | 7.25 |
| 50 (5-20 trained) | 9.14 |
Here is the list of changes I did to the original paper description
-
Convex Hull decoding: The paper states that the decoding process for the convex hull task was unconstrained. I added some conditions to the decoding process for convex hull, see the details here.
-
Optimizer and Scheduler: Changed SGD to Adam, learn rate from 1.0 to 1e-3 (inline with what's recommended for Adam) and added a learn rate scheduler (exponential decay). The optimizer largely improved and sped up convergence, and the scheduler helped with reducing noise in the final steps of training.
Some were useful, some weren't