I can not reproduce the experimental results by using the .pt file in the pretrained folder. about omni-vrp HOT 3 CLOSED

I just wanna check the specific model of pre-trained weights. is it correct?

maml-checkpoint-{epoch}-batch-no-track.pt : ours-SO settings. Omni-VRP w/ fully second order. maml-fomaml-checkpoint-{epoch}-batch-no-track.pt : ours settings. Omni-VRP w/ early stopping second order. pomo-checkpoint-{epoch}-batch.pt : POMO settings. POMO trained in the meta-training way. reptile-checkpoint-{epoch}-instance.pt : Meta-POMO settings. Manchanda et al., 2022 work.

Exactly.

from omni-vrp.

Hi, the .pt file is not the (meta-) pretrained model. It is the model pretrained on uniformly distributed data following the POMO source code. It is provided since the meta-training could also apply to the pretrained model.

For reference, I have uploaded the meta-pretrained models at 50K and 250K iteration. I found them in my log files and did not check the results (but should be around ~2%). For fully reproduce the results, please run train.py.

from omni-vrp.

I just wanna check the specific model of pre-trained weights. is it correct?

maml-checkpoint-{epoch}-batch-no-track.pt : ours-SO settings. Omni-VRP w/ fully second order.
maml-fomaml-checkpoint-{epoch}-batch-no-track.pt : ours settings. Omni-VRP w/ early stopping second order.
pomo-checkpoint-{epoch}-batch.pt : POMO settings. POMO trained in the meta-training way.
reptile-checkpoint-{epoch}-instance.pt : Meta-POMO settings. Manchanda et al., 2022 work.

from omni-vrp.