Something is wrong in revMHA Encoder, which project the nodes coordinates to the weird embedding like [nan,nan,nan,nan,...]

[bxx-seu]

Dear author, I ran the code of Pointerformer locally and tested its performance to TSP tasks. However, I found that the model could not correctly solve the tasks for some instances, and obtained inexplicable final solutions such as [0,0,0,0,0,0].

By debug, found that something is wrong in revMHA Encoder, which cause the node embeddings projected from the model input becoming the weird embedding, like [nan,nan,nan,nan,...]. Abviously, the solution must be wrong.

For config.yaml, I only changed the "precision" from 16 to 64 because of "RuntimeError: expected scalar type Double but found Half".

I attach a testset for verification, which include four instances and the first can be solved normally and the rest three instances can not be solved.
testset.txt

[neo-pan]

The issue you encountered is due to a known problem where the distribution of the training and test data differs. We have implemented adjustments in the code to handle out-of-distribution data through a simple scaling mechanism. You can control this setting using the real_data flag. Setting this flag to True should provide the correct result for your case.

https://github.com/Learning4Optimization-HUST/Pointerformer/blob/1e79b6ac65e2cdb3c5d91b5dd46b8b1731a2557e/eval.py#L210C1-L218C27

Apologies for the delayed response, as I do not check this repo frequently. If you have any further questions or need assistance, feel free to reach out!