Understanding property_embeddings vs property_embeddings_adapt

[luisbro]

Hello, I would like to ask about the difference of the variables property_embeddings and property_embeddings_adapt as found in the code and the yaml config files. I could not really find what property_embeddings is used for, but from issues #58, #128 I would assume the following:

• property_embeddings: Training a model from scratch, conditioned on the specified properties
• property_embeddings_adapt: Fine-tuning a model to condition it on the specified properties

Is this accurate? If so, it seems to me like all models provided here use the fine tuning via property_embeddings_adapt, so what made you decide to keep the option of training from scratch via property_embeddings?

Thanks (:

---

As a side note, the docstring on the GemNetTDenoiser init seems a bit outdated, quoting an incorrect default for hidden dims and not matching all of the current variables, like property_embeddings and property_embeddings_adapt (property_embeddings_adapt is unused here). I could try to update it, if you won't find the time and would like me to.

mattergen/mattergen/denoiser.py

Lines 173 to 195 in 5bb2b39

	def __init__(
	self,
	gemnet: nn.Module,
	hidden_dim: int = 512,
	denoise_atom_types: bool = True,
	atom_type_diffusion: str = [
	"mask",
	"uniform",
	][0],
	property_embeddings: torch.nn.ModuleDict | None = None,
	property_embeddings_adapt: torch.nn.ModuleDict | None = None,
	element_mask_func: Callable | None = None,
	**kwargs,
	):
	"""Construct a GemNetTDenoiser object.
	Args:
	gemnet: a GNN module
	hidden_dim (int, optional): Number of hidden dimensions in the GemNet. Defaults to 128.
	denoise_atom_types (bool, optional): Whether to denoise the atom types. Defaults to False.
	atom_type_diffusion (str, optional): Which type of atom type diffusion to use. Defaults to "mask".
	condition_on (Optional[List[str]], optional): Which aspects of the data to condition on. Strings must be in ["property", "chemical_system"]. If None (default), condition on ["chemical_system"].
	"""

[danielzuegner]

Hi,

you're correct about the difference in what property_embeddings and property_embeddings_adapt are. We kept the option of training on a property from scratch because it may be useful for some users.

If you feel like updating the docstring, feel free to issue a PR and we'll merge it after some review.

Best,

Daniel

[luisbro]

Hey, thanks for your help (:

I created a PR for it #134