Notes

Installation

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

Training

The paper reports results using the Rollout algorithm (Section 4.2), noting PPO performs slightly less well. Rollout checkpoints are directly compatible with the evaluator (no key conversion needed). n_nodes is hardcoded to 21 in both trainers; there is no CLI argument.

Rollout (recommended)

CUDA_VISIBLE_DEVICES=1 python -m VRP.VRP_Rollout_train

Saves each epoch to Vrp-21-GAT/rollout/{epoch}/actor.pt.

PPO

CUDA_VISIBLE_DEVICES=1 python -m VRP.PPO_train

Saves each epoch to vrp-21-GAT/20201125/{epoch}/actor.pt. PPO checkpoints need key conversion before use with the evaluator (see below).

Evaluation

Need folder "./trained".

Using a self-trained checkpoint

PPO_train.py saves Actor_critic.state_dict(), whose keys have an actor. prefix incompatible with the plain Model expected by test_vrp.py. Convert before placing the checkpoint in trained/<n>/actor.pt:

python scripts/fix_actor_keys.py vrp-21-GAT/20201125/99/actor.pt trained/21.mine/actor.pt

CUDA_VISIBLE_DEVICES=1 python -m VRP.test_vrp.py