Incompatible Pre-trained Model being pulled

[shubharajkharel]

Error

• Loading pre-trained model while featurizing raises this BaseException exception.

Cause

• load_model is not pulling the version of pre-trained model from matgl:v0.8.5 being used in cresendo but a newer model which includes fixes not compatible with matgl:v0.8.5

Temporary Fix

• Replace the pre-trained model (default: M3GNet-MP-2021.2.8-PES) files cached in directory ~/.cache/matgl/ with the older pre-trained version.

[matthewcarbone]

Can you post the full stack trace? This is really surprising. The default model should be version-locked so I'm not sure why this isn't working.

It also seems like this might be a bug with matgl's deployment of their model, no?

[shubharajkharel]

This is really surprising.

I am too..

It also seems like this might be a bug with matgl's deployment of their model, no?

I agree, this is caused by undetected bug in matgl.

Call

featurize_material(Structure.from_file(poscar_path))

Stack Trace

Traceback (most recent call last):
  File "/Users/bnl/Downloads/dev/aimm/ai_for_xas/temp.py", line 10, in <module>
    from Crescendo.crescendo.extern.m3gnet._featurizer import featurize_material
  File "/Users/bnl/Downloads/dev/aimm/ai_for_xas/Crescendo/crescendo/extern/m3gnet/__init__.py", line 1, in <module>
    from ._featurizer import featurize_material  # noqa
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/bnl/Downloads/dev/aimm/ai_for_xas/Crescendo/crescendo/extern/m3gnet/_featurizer.py", line 58, in <module>
    def featurize_material(structure, model=_load_default_featurizer()):
                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/bnl/Downloads/dev/aimm/ai_for_xas/Crescendo/crescendo/extern/m3gnet/_featurizer.py", line 53, in _load_default_featurizer
    model = load_model("M3GNet-MP-2021.2.8-PES").model
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/bnl/mambaforge/envs/aimm/lib/python3.11/site-packages/matgl/utils/io.py", line 213, in load_model
    raise ValueError(
ValueError: Bad serialized model or bad model name. It is possible that you have an older model cached. Please clear your cache by running `python -c "import matgl; matgl.clear_cache()"`

[matthewcarbone]

Have you tried clearing the cache?

[shubharajkharel]

Have you tried clearing the cache?

Of course. Error message itself mentions it!

But it doesn't work because for reason I outlined above. So temporary solution involves replacing file in cache directory, which just copy pastes the working files into it.

def fix_m3gnet_version(
    model_name="M3GNet-MP-2021.2.8-PES",
    cache_dir=None,
):
    if cache_dir is None:
        cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "matgl", model_name)
    valid_model_dir = os.path.join("scripts", model_name)
    if not os.path.exists(valid_model_dir):
        raise FileNotFoundError(f"Model directory {valid_model_dir} does not exist. ")
    try:
        shutil.copytree(valid_model_dir, cache_dir)
    except FileExistsError:
        warnings.warn(f"Model directory {cache_dir} already exists. ")
        warnings.warn("Overwriting existing model. ")
        shutil.rmtree(cache_dir)
        shutil.copytree(valid_model_dir, cache_dir)
    except Exception as e:
        raise e

Also note:

• This issue is resolved in the current version. But updating it to current version creates problem in other places of current code in Cresendo starting with how Structure2Graph is implemented in newer version of matgl.
• The fix is not about change of model, but changes in model.json and few other files.
• Relevant changes after 0.8.5: Bug report, Pull request

[matthewcarbone]

Ok thanks for digging. Long story short, I would just use whatever featurizer that is available (probably preferable to your hack, which is nice by the way). It's not so important, in my opinion. As long as it comes from some "universal" model it will prove our point. Also that branch of Crescendo is now kinda stale and outdated. I'm not really working on it/testing it anymore.