Skip to content

Latest commit

 

History

History
82 lines (68 loc) · 5.24 KB

README.md

File metadata and controls

82 lines (68 loc) · 5.24 KB
Raw

GNN baseline (supervised)

This directory contains all the code to train a ML baseline: a 2-layer Graph Neural Network (GNN) architecture with GraphConv convolutions. The model is trained using supervised learning, where training data is generated using the oracle solver. Hyperparameter tuning has not been performed extensively, therefore much performance can be gained by a proper hyperparameter optimization scheme. You may build on this, for example by optimizing the hyperparameters and experimenting with different types of convolutions and neural architectures.

Install requirements

This baseline requires PyTorch and PyTorch Geometric. Install using:

# If you want to use GPU, do not use pip but [install PyTorch](https://pytorch.org/get-started/locally/) manually and run below commands instead (without #)
# pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-$(python -c "import torch; print(torch.__version__)").html
# pip install torch-geometric
# Run from root of repository (install CPU version using pip)
pip install -r baselines/supervised/requirements.txt

Generate data, train and test a model

Quick example to generate 10 data points using two instances with 5 seeds each, use it to train a model and test the model within the controller (using short time limits and few epochs for a quick example):

# Run from root of repository
python baselines/supervised/generate_training_data.py  --instance instances/ORTEC-VRPTW-ASYM-0bdff870-d1-n458-k35.txt --epoch_tlim 2 --oracle_tlim 5 --verbose
python baselines/supervised/generate_training_data.py  --instance instances/ORTEC-VRPTW-ASYM-08d8e660-d1-n460-k42.txt --epoch_tlim 2 --oracle_tlim 5 --verbose
python baselines/supervised/process_training_data.py
python baselines/supervised/train.py --num_epochs 10
python controller.py --instance instances/ORTEC-VRPTW-ASYM-0bdff870-d1-n458-k35.txt --epoch_tlim 5 -- python solver.py --verbose --strategy supervised --model_path baselines/supervised/ckpt

Generate training data

In order to generate training data, you need to first execute the script python baselines/supervised/generate_training_data.py from the root directory. This script requires the name of the instance which will be used to generate data from. The command line interface is equivalent to that of solver.py. Therefore, participants may make use of the utility script run_parallel_solver.sh to generate training data for all instances as well.

Running generate_training_data.py will create a directory data, which is supposed to collect all the data from all the instances. After that, you can run python baselines/supervised/process_training_data.py, which aggregates and transform the data to be ready to be used by the train.py script in this directory. The output of this script is one file for the input features (data.X.pkl) and one file for the ground truth labels (data.Y.pkl).

Download training data

Ready to use training data can be downloaded here. Make sure to put the files in the folder baselines/supervised/data.

Train the model

  python baselines/supervised/train.py

This script will dump model's checkpoints to the directory ckpt located in the root directory.

Inference

To use the trained model as a strategy, you just need to pass --strategy supervised to either solver.py or controller.py along with the checkpoint path with --model_path <CKPT_DIR>.

Pretrained model

We provide a pretrained GNN baseline to ease reproducibility of the performance. The checkpoint is saved at baselines/supervised/pretrained and can be tested by passing --strategy supervised --model_path ./baselines/supervised/pretrained to solver.py.

The current model takes as input a complete graph representing the current VRP epoch instance. Node features comprise:

  • A binary feature indicating whether the node is the depot.
  • Normalized node demands.
  • A binary feature indicating whether the node must be dispatched in current epoch. The cost of driving along the edge direction is used as the only edge feature.

Participants are free to enrich node and edge features with anything valuable to ease learning and performance of the neural network.

Test and create submission

Make sure your solver can run through baselines/supervised/run.sh, then use the baselines/supervised/create_submission.sh script to create a submission (you can omit any training code but include the trained model and necessary files!).

python controller.py --instance instances/ORTEC-VRPTW-ASYM-0bdff870-d1-n458-k35.txt --epoch_tlim 5 -- baselines/supervised/run.sh
baselines/supervised/create_submission.sh
# Test the submission to see if all dependencies where packed correctly
cd submissions/tmp
./install.sh && python ../../controller.py --instance ../../instances/ORTEC-VRPTW-ASYM-0bdff870-d1-n458-k35.txt --epoch_tlim 5 -- ./run.sh
cd ../..

Hyperparameters & Performances

Hyperparameter Value
Learning Rate 0.01
Weight Decay 0.005
Hidden Neurons 64
Train Validation Test
~66% ~68% ~68%