Pytorch implementation of encode-attend-navigate, a Deep Reinforcement Learning based TSP solver.

You can leverage the free GPU on Colab to train this model. Just run this notebook :

Clone the repository :

git clone https://github.com/astariul/encode-attend-navigate-pytorch.git
cd encode-attend-navigate-pytorch

Install dependencies :

pip install -r requirements.txt

Run the code :

python main.py

You can specify your own configuration file :

python main.py config=my_conf.yaml

Or directly modify parameters from the command line :

python main.py lr=0.002 max_len=100 batch_size=64

I ran the code with the following command line :

python main.py enc_stacks=1 lr=0.0002 p_dropout=0.1

On Colab, with a Tesla T4 GPU, it tooks 1h 46m for the training to complete.

Here is the training curves :

After training, here is a few example of path generated :

This code is a direct translation of the official TF 1.x implementation, by @MichelDeudon.

Please refer to their README for additional details.

To ensure the Pytorch implementation produces the same results as the original implementation, I compared the outputs of each layer given the same inputs and check if they are the same.

You can find (and run) these tests on this Colab notebook :