yd-kwon / matnet

codes for the paper "Matrix Encoding Networks for Neural Combinatorial Optimization"

MatNet

This repository provides a reference implementation of MatNet and saved trained models as described in the paper:

Matrix Encoding Networks for Neural Combinatorial Optimization
(NeurIPS 2021, accepted)
https://arxiv.org/abs/2106.11113

The code is written using Pytorch.

Getting Started

We provide codes for two CO (combinatorial optimization) problems:

• Asymmetric Traveling Salesman Problem (ATSP)
• Flexible Flow Shop Problem (FFSP)

Basic Usage

For both ATSP_MatNet and FFSP_MatNet,

i. To train a model,

python3 train.py

train.py contains parameters you can modify.
At the moment, it is set to train N=20 problems.

ii. To test a model,

python3  test.py

You can specify the model as a parameter contained in test.py.
At the moment, it is set to use the saved model (N=20) we have provided (in "result" folder), but you can easily use the one you have trained from running train.py.

To test for the N=50 model, make sure that saved problem files exist in the path (see below for Datasets). Also, modify test.py so that

• all "20"'s are changed to "50"
• "path" and "epoch" in the "tester_params" are correctly pointing to the saved model
• test batch size is decreased (by a factor of something like 4)

Saved Trained Models

Trained models for ATSP are provided with the codes.
However, the sizes of the trained model files for FFSP-MatNet are too large to upload here. So we provide them as links below.

• MatNet_saved_models_FFSP.zip (size: 358MB) <-- a zip file containing all three models
  • MatNet_saved_models_N100 (size: 119MB)
  • MatNet_saved_models_N50 (size: 119MB)
  • MatNet_saved_models_N20 (size: 119MB)

Datatsets

Test datasets for larger N (N=50, N=100) problems are given as links below.
Download the dataset and add them to the "data" directories under ATSP and FFSP folders.

• MatNet_dataset_ATSP (size: 374MB)
• MatNet_dataset_FFSP (size: 19.6MB)

Used Libraries

python v3.7.6
torch==1.7.0

matnet's Issues

multi-GPU implementation

In your paper, you mentioned that for n=50/100 large scale data, you adopted a multi-GPU implementation, I wonder if it is possible to provide a multi-GPU version of the code

I want ask a question?

Sorry because i dont understand the code and I need your help ! I have two question

• The first question: I understand that the score and aug_score is the tour length. But went i test it with my own dataset, i change the scale so the ans is different. So i want to ask that the train scale and test scale must be the same ?
• The second question: I know that max_pomo_reward.max(dim=0) is the best tour and I dont know how to get it, and I know self.env.slected_node_list is the tour but how i can get it from the best tour ?
  TKS!!

I have a question

May I ask if the number of machines in each stage of the FFSP problem needs to be consistent? If it can be a heterogeneous quantity, how should pomo_size be set?

I have some question !

What can i do with dataset like tsp with time windown thks

The model file is not exist

Hi, it seems that the trained model saved in results/saved_atsp_model is not available now. I tried to get the model from git lfs, but it shows the model does not exist on the server. Could you please solve this problem? Thanks very much.

I have a question?

max_aug_pomo_reward, _ = max_pomo_reward.max(dim=0) # get best results from augmentation
# shape: (batch,)
aug_score = -max_aug_pomo_reward.float().mean() # negative sign to make positive value

I want to ask why you use .mean() but we want to get the best result

Hope you can help me

If I met the Multiple data matrices, what should I do， May I do like this ： scores = torch.stack((dot_product_score, cost_mat1, cost_mat2, cost_mat3 ), dim=4).
There may also be papers that I didn't fully understand, forgive me for being a newbie

Ask some questions

Have you ever tried to solve the dynamic vehicle scheduling problem with this model？

I have a question!

I have a question about job scheduling problem. My my problem is very complicated not like job shop so what if i change the problems define as meta heuristic algorithms as sample for model ?

I have a question

How I modify the code for symmetric TSP