[Feat]: Direct property predictor

[simplecaryu]

Contact Details

No response

Feature Description

Enable fine-tuning of the end-to-end model for direct property prediction, covering properties such as band gap or dielectric constant, as mentioned in the paper

Motivation

The paper mentions direct property prediction features, but it seems they aren't available in the current repository. Are there any plans to make these features public? It is also related to the problems discussed in issue #108.

Proposed Solution

No response

Contribution Interest

• I'm interested in potentially implementing this feature
• I can provide guidance, but cannot implement
• I'm just suggesting the feature for the community

Code of Conduct

• I agree to follow the project's Code of Conduct

[luisbro]

I would also be interested in seeing some code for this being released. Some fragments from an existing implementation can still be seen, but making them more usable and providing some more details on the architecture of the finetune heads and how they were trained, would be great.

mattersim/src/mattersim/forcefield/potential.py

Lines 158 to 184 in 301c11b

	def finetune_mode(
	self,
	finetune_layers: int = -1,
	finetune_head: nn.Module = None,
	reset_head_for_finetune: bool = False,
	finetune_task_mean: float = 0.0,
	finetune_task_std: float = 1.0,
	use_finetune_label_loss: bool = False,
	):
	"""
	Set the model to fine-tuning mode
	finetune_layers: the layer to finetune, former layers will be frozen
	if -1, all layers will be finetuned
	finetune_head: the head to finetune
	reset_head_for_finetune: whether to reset the original head
	"""
	if self.model_name not in ["graphormer", "geomformer"]:
	logger.warning("Only graphormer and geomformer support freezing layers")
	return
	self.model.finetune_mode = True
	if finetune_head is None:
	logger.info("No finetune head is provided, using the original energy head")
	self.model.finetune_head = finetune_head
	self.model.finetune_task_mean = finetune_task_mean
	self.model.finetune_task_std = finetune_task_std
	self.freeze_reset_model(finetune_layers, reset_head_for_finetune)
	self.use_finetune_label_loss = use_finetune_label_loss