Pointer Networks in Tensorflow
TensorFlow implementation of Pointer Networks.
Support multithreaded data pipelines to reduce I/O latency.
Requirements
- Python 2.7
- requests
- tqdm
- TensorFlow 0.12.1
Usage
To train a model:
$ python main.py --task=tsp --max_data_length=20 --hidden_dim=512 # download dataset used in the paper
$ python main.py --task=tsp --max_data_length=10 --hidden_dim=128 # generate dataset itself
To train a model:
$ python main.py
$ tensorboard --logdir=logs --host=0.0.0.0
To test a model:
$ python main.py --is_train=False
Results
Train/Test loss of max_data_length=10 after 24,000 steps:
$ python main.py --task=tsp --max_data_length=10 --hidden_dim=128
$ ...
$ 2017-01-18 17:17:48,054:INFO::
$ 2017-01-18 17:17:48,054:INFO::test loss: 0.00149279157631
$ 2017-01-18 17:17:48,055:INFO::test pred: [1 6 2 5 4 3 0 0 0 0 0]
$ 2017-01-18 17:17:48,057:INFO::test true: [1 6 2 5 4 3 0 0 0 0 0] (True)
$ 2017-01-18 17:17:48,059:INFO::test pred: [1 3 8 4 7 5 2 6 0 0 0]
$ 2017-01-18 17:17:48,059:INFO::test true: [1 3 8 4 7 5 2 6 0 0 0] (True)
$ 2017-01-18 17:17:48,058:INFO::test pred: [ 1 8 3 7 9 5 6 4 2 10 0]
$ 2017-01-18 17:17:48,058:INFO::test true: [ 1 8 3 7 9 5 6 4 2 10 0] (True)
Train/Test loss of max_data_length=20 after 16,000 steps:
$ python main.py --task=tsp --max_data_length=20 --hidden_dim=512
$ ...
$ 2017-01-18 17:30:44,738:INFO::
$ 2017-01-18 17:30:44,739:INFO::test loss: 0.0501566007733
$ 2017-01-18 17:30:44,739:INFO::test pred: [ 1 3 16 18 6 2 7 4 5 15 10 8 13 11 17 20 19 14 9 12 0]
$ 2017-01-18 17:30:44,740:INFO::test true: [ 1 3 16 18 6 2 7 4 5 15 10 8 13 11 17 20 19 14 9 12 0] (True)
$ 2017-01-18 17:30:44,741:INFO::test pred: [ 1 4 12 8 2 17 20 3 18 14 16 11 13 15 7 10 9 6 19 5 0]
$ 2017-01-18 17:30:44,741:INFO::test true: [ 1 4 12 8 2 17 20 3 18 14 16 11 13 15 7 10 9 6 19 5 0] (True)
$ 2017-01-18 17:30:44,742:INFO::test pred: [ 1 8 16 20 19 14 18 2 9 12 15 11 3 13 5 6 7 4 17 10 0]
$ 2017-01-18 17:30:44,742:INFO::test true: [ 1 8 16 20 19 14 18 2 9 12 15 11 3 13 5 6 7 4 17 10 0] (True)
Author
Taehoon Kim / @carpedm20