hatemhelal / mace Goto Github PK

The 3BPA dataset has a number of different training / test splits. Table 2 from the MACE paper reports test errors after training MACE variants on the 300K training set.

One possible sketch:

• add a new optional command line argument --additional_tests that takes one or more xyz file paths to test
• after reaching max_num_epochs the test error is collected for every additional test path
• when using wandb, the test errors should be logged for later comparison

It would be useful to have the final checkpoint of a training run logged to wandb to make it easier to transfer between different machines.

Is your feature request related to a problem? Please describe.
I don't think we track these at the moment

Describe the solution you'd like
We can do it using register_backward_hook, presumably

Describe alternatives you've considered
We could just not do it and hope for the best idk

Additional context
No

@callumm-graphcore added logging of the initial test loss. I think we can use this to convince ourselves that the models are being initialised the same irrespective of the choice for the default data type.