Learning to Solve Many-to-Many Pickup and Delivery Problems with Multi-head Heterogeneous Attention

Implementation for the paper “Learning to Solve Many-to-Many Pickup and Delivery Problems with Multi-head Heterogeneous Attention”.

Requirements

• Python 3.10
• Training requires NVIDIA GPU + CUDA (uses NCCL + torchrun / DDP)

Installation

conda create -n hap python=3.10 -y
conda activate hap

pip install -r requirements.txt

Usage

Generate validation data

python generate_data.py --name validation --graph_sizes 20 --seed 12345

Generate testing data

python gen_test_data.py

Train (single GPU)

graph_node=20
validation_data_seed=12345

torchrun --nproc_per_node=1 run.py \
  --graph_size $graph_node \
  --run_name "n${graph_node}_main" \
  --val_dataset "data/pdp/pdp${graph_node}_validation_seed${validation_data_seed}.pkl"

Train (multi-GPU, example 4 GPUs)

torchrun --nproc_per_node=4 run.py \
  --graph_size 20 \
  --run_name "n20_main" \
  --val_dataset "data/pdp/pdp20_validation_seed12345.pkl"

Evaluate

1. Edit the config of eval_main.py:

• Set checkpoint paths in models = {...}
• Set dataset directory in data_files = ...
• Set output directory in out_dir = ...

2. Run:

python -u eval_main.py

Acknowledgement

Thanks to Wouter Kool et al. and Jingwen Li et al. for getting me started with the code.