Skipping scheduler for EquiformerV2 with OC22

[sedaoturak]

Hello!
There was an issue regarding skipping the scheduler of Equiformer in case of OCPCalculater use (issue #541 ). Is it possible to do the same thing for EquiformerV2 model (checkpoint ) trained with OC22?

[mshuaibii]

Hmm - the issue you pointed to should address any instances of EqV2. Can you share a minimal example that I can use to reproduce the error you're running into.

[sedaoturak]

Sure. I came across this error when I tried to create the OCPCalculator object:
(I work on google colab and skip the other installation code snippets in the notebook)

!wget https://dl.fbaipublicfiles.com/opencatalystproject/models/2023_10/oc22/s2ef/eq2_121M_e4_f100_oc22_s2ef.pt

from ocpmodels.common.relaxation.ase_utils import OCPCalculator
# checkpoint_path = "gnoc_oc22_oc20_all_s2ef.pt"
checkpoint_path = '/content/ocp/eq2_121M_e4_f100_oc22_s2ef.pt'
ocp_calculator = OCPCalculator(checkpoint_path=checkpoint_path)

---

TypeError Traceback (most recent call last)
in <cell line: 5>()
3 # checkpoint_path = "gnoc_oc22_oc20_all_s2ef.pt"
4 checkpoint_path = '/content/ocp/eq2_121M_e4_f100_oc22_s2ef.pt'
----> 5 ocp_calculator = OCPCalculator(checkpoint_path=checkpoint_path)

5 frames
/content/ocp/ocpmodels/modules/scheduler.py in init(self, optimizer, config)
34 self.scheduler = getattr(lr_scheduler, self.scheduler_type)
35 scheduler_args = self.filter_kwargs(config)
---> 36 self.scheduler = self.scheduler(optimizer, **scheduler_args)
37
38 def step(self, metrics=None, epoch=None) -> None:

TypeError: LambdaLR.init() missing 1 required positional argument: 'lr_lambda'

[mshuaibii]

Sorry for the delay on this, I indeed was able to reproduce this error.

I have uploaded the checkpoint at the same location, you should be able to redownload the file and try again.

[sedaoturak]

Hi, I've tried again and got the same error. I've also converted the code to fairchem. I don't know if I do something wrong:

ocp_calculator = OCPCalculator(
    model_name="EquiformerV2-lE4-lF100-S2EFS-OC22",
    local_cache="pretrained_models",
    cpu=False,
)


WARNING:root:Detected old config, converting to new format. Consider updating to avoid potential incompatibilities.
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
[<ipython-input-12-1982fd4316ff>](https://localhost:8080/#) in <cell line: 8>()
      6 # ocp_calculator = OCPCalculator(checkpoint_path=checkpoint_path)
      7 
----> 8 ocp_calculator = OCPCalculator(
      9     model_name="EquiformerV2-lE4-lF100-S2EFS-OC22",
     10     local_cache="pretrained_models",

5 frames
[/usr/local/lib/python3.10/dist-packages/fairchem/core/modules/scheduler.py](https://localhost:8080/#) in __init__(self, optimizer, config)
     39             self.scheduler = getattr(lr_scheduler, self.scheduler_type)
     40             scheduler_args = self.filter_kwargs(config)
---> 41             self.scheduler = self.scheduler(optimizer, **scheduler_args)
     42 
     43     def step(self, metrics=None, epoch=None) -> None:

TypeError: LambdaLR.__init__() missing 1 required positional argument: 'lr_lambda'

[mshuaibii]

Can you run md5sum eq2_121M_e4_f100_oc22_s2ef.pt on the checkpoint that downloaded. I ran your exact commands with no issue.

[sedaoturak]

It seems it works. Thank you!