Bug for fine-tuning the OMat24 model

[jinlhr542]

I am trying to fine-tuning the checkpoint_sevennet_mf_ompa.pth model:

import sevenn.util as util
model, config = util.model_from_checkpoint('checkpoint_sevennet_mf_ompa.pth')
cutoff = config['cutoff'] 
dataset = SevenNetGraphDataset(cutoff=cutoff, root=working_dir, files=dataset_files, processed_name='train.pt')

from sevenn.train.trainer import Trainer
import torch.optim.lr_scheduler as scheduler

trainer = Trainer.from_config(model, config)

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[6], line 4
      1 from sevenn.train.trainer import Trainer
      2 import torch.optim.lr_scheduler as scheduler
----> 4 trainer = Trainer.from_config(model, config)
      6 # We have energy, force, stress loss function, which used to train 7net-0.
      7 # We will use it as it is, with loss weight: 1.0, 1.0, and 0.01 for energy, force, and stress, respectively.
      8 print(trainer.loss_functions)

File [~/miniconda3/envs/atomate2/lib/python3.12/site-packages/sevenn/train/trainer.py:88](http://localhost:3416/lab/tree/DRX/miniconda3/envs/atomate2/lib/python3.12/site-packages/sevenn/train/trainer.py#line=87), in Trainer.from_config(model, config)
     84 @staticmethod
     85 def from_config(model: torch.nn.Module, config: Dict[str, Any]) -> 'Trainer':
     86     trainer = Trainer(
     87         model,
---> 88         loss_functions=get_loss_functions_from_config(config),
     89         optimizer_cls=optim_dict[config.get(KEY.OPTIMIZER, 'adam').lower()],
     90         optimizer_args=config.get(KEY.OPTIM_PARAM, {}),
     91         scheduler_cls=scheduler_dict[
     92             config.get(KEY.SCHEDULER, 'exponentiallr').lower()
     93         ],
     94         scheduler_args=config.get(KEY.SCHEDULER_PARAM, {}),
     95         device=config.get(KEY.DEVICE, 'auto'),
     96         distributed=config.get(KEY.IS_DDP, False),
     97         distributed_backend=config.get(KEY.DDP_BACKEND, 'nccl'),
     98     )
     99     return trainer

File [~/miniconda3/envs/atomate2/lib/python3.12/site-packages/sevenn/train/loss.py:211](http://localhost:3416/lab/tree/DRX/miniconda3/envs/atomate2/lib/python3.12/site-packages/sevenn/train/loss.py#line=210), in get_loss_functions_from_config(config)
    207 from sevenn.train.optim import loss_dict
    209 loss_functions = []  # list of tuples (loss_definition, weight)
--> 211 loss = loss_dict[config[KEY.LOSS].lower()]
    212 loss_param = config.get(KEY.LOSS_PARAM, {})
    214 use_weight = config.get(KEY.USE_WEIGHT, False)

AttributeError: 'dict' object has no attribute 'lower'

[YutackPark]

ompa is trained on an internal repository that does not fully compatible with public SevenNet. Therefore, current python interface need some manual intervention to make format compatible. Instead, you could try fine-tuning via CLI interface

sevenn preset mf_ompa_fine_tune > input.yaml

This will yield working input.yaml file for fine-tuning the ompa model.

• Please note that I don't have a near plan to make the fully functioning python interface for the latest models as we have found that training code is extremly hard to maintain with consistency.