Code for running GLOP

[ujjwaldasari10]

Can you please give an example code for training and inference of GLOP for lets say TSP of size 10K?

[fedebotu]

Hi @ujjwaldasari10, we did not merge the code for GLOP yet since it is not refactored so far

CC @Furffico @henry-yeh

[DasariShreeUjjwal]

@fedebotu, @cbhua I am adding a couple of more doubts here. 1) Is it possible to save a trained EAS model and apply the model to a newer test instance other than the one it was saved on? 2) In the original DEEPACO paper the models are run for some 'evaluations' and not epochs. Can you please let me know the equivalent implementation in RL4CO since it's not trivial to me, going through the code base? Thanks

[Furffico]

Hi @DasariShreeUjjwal 👋🏻, thank you for your interest in our work!
With regard to question 2:

In the original DEEPACO paper the models are run for some 'evaluations' and not epochs. Can you please let me know the equivalent implementation in RL4CO since it's not trivial to me, going through the code base?

The "# of evaluations" in the DeepACO paper refers to the number of evaluating the performance (route length) of solutions during an ACO run on a single problem instance. However, the "epochs" in rl4co is the number of total iterations for the whole training loop, which is more similar to the Total training instances in Table 7. This can be expressed as Total training instances = epochs * train_data_size.

For the former concept, we have evaluations = ACO_iterations * n_ants, and you may set these two parameters accordingly like this:

from rl4co.models import DeepACO
model = DeepACO(
    ...,
    policy_kwargs = dict(n_ants=20, n_iterations=10)
)

[Furffico]

Can you please give an example code for training and inference of GLOP for lets say TSP of size 10K?

BTW, I'm currently working on reimplementing GLOP in RL4CO, and it's almost complete. You might check the codebase here: https://github.com/Furffico/rl4co/tree/dev-glop/rl4co/models/zoo/glop