PtrNet

Chainer implementation of ptrnet for TSP.
The dataset is not available here so please it prepare by yourself. Discription is in /nets/base_network.py.

how to use

step1.

Install the chainer.

pip install chainer=='1.24.0'

step2.

Create the dataset as discribed in base_network.py.
I used 1048576 tours for train data.

step3.

Run the code for training.

python main.py

step4.

After training 10 epochs, run the code for testing

python main.py --mode test --load ptrnet 10 --beam_width 10

Beam search will be done if you assign the beam_width > 1.

result

The original paper said the average tour length predicted from ptrnets is 2.88 compared with its ground truth length 2.87, and actually I got 2.88.