Errors with ODAC23 EquiformerV2 models

[andreadaru]

Hello!
I found some errors when I try to run the new ODAC23 EquiformerV2 models. The same code with other ODAC23 models works.

With EquiformerV2-S2EF-ODAC the error is:

 File "/fairchem/core/common/relaxation/ase_utils.py", line 184, in __init__
    self.trainer = registry.get_trainer_class(config["trainer"])(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/fairchem/core/trainers/ocp_trainer.py", line 96, in __init__
    super().__init__(
  File "/fairchem/core/trainers/base_trainer.py", line 182, in __init__
    self.load()
  File "/fairchem/core/trainers/base_trainer.py", line 209, in load
    self.load_extras()
  File "/fairchem/core/trainers/base_trainer.py", line 597, in load_extras
    self.scheduler = LRScheduler(self.optimizer, self.config["optim"])
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/fairchem/core/modules/scheduler.py", line 41, in __init__
    self.scheduler = self.scheduler(optimizer, **scheduler_args)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: LambdaLR.__init__() missing 1 required positional argument: 'lr_lambda'

With EquiformerV2-Large-S2EF-ODAC the error is:

   calc = OCPCalculator(
           ^^^^^^^^^^^^^^
  File "/fairchem/core/common/relaxation/ase_utils.py", line 184, in __init__
    self.trainer = registry.get_trainer_class(config["trainer"])(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/fairchem/core/trainers/ocp_trainer.py", line 96, in __init__
    super().__init__(
  File "/fairchem/core/trainers/base_trainer.py", line 182, in __init__
    self.load()
  File "/fairchem/core/trainers/base_trainer.py", line 206, in load
    self.load_model()
  File "/fairchem/core/trainers/base_trainer.py", line 418, in load_model
    self.model = registry.get_model_class(self.config["model"])(
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: EquiformerV2_OC20.__init__() got an unexpected keyword argument 'norm_scale_nodes'

Any help?

[mwzum]

Hello,

First of all, thank you for the incredible work on this project.

Given that the OC models work perfectly, it seems the problem is specific to the ODAC-related models. Could this be a compatibility issue with the configuration format or something specific to the ODAC model settings?

I have attempted to add/remove some keys directly into the configuration, but this leads to other issues arising. viz.,
KeyError: 'dataset' in the config["task"]["dataset"] key.

I also noticed @janiceblue and @mshuaibii updated the configs to use ocp2.0 format (#653)

@anuroopsriram @abhshkdz: Any guidance or updates on resolving these configuration issues would be greatly appreciated.

Thank you for your time and assistance!

[mshuaibii]

@anuroopsriram this looks like ODAC used an older version of EqV2. The scheduler issue is something you'll also need to manually update the checkpoints, similar to #660. Can you please take a look.

[komorisu]

I greatly appreciate your excellent research.

Since the EquiformerV2-IE4-IF100-S2EFS-OC22 trained with OC22 worked well, I agree with your opinion. Could you please release an ODAC using the new version of EqV2 in the future?

Thank you for your assistance.

[zulissimeta]

I added a unit test here for the S2EF/IS2RE and ODAC configs. There's still some problems with the IS2RE configs:
#766

@andreadaru Are there any other configs that are breaking that we should take a look at?

[github-actions]

This issue has been marked as stale because it has been open for 30 days with no activity.

[github-actions]

This issue has been marked as stale because it has been open for 30 days with no activity.