Is fine-tuning necessary for property conditioned structures generation?

[Mofahdi]

Do I have to fine-tune a base model to generate new structures with a property condition?
let us say that I want to train a model to get the band gap. From my understanding, I cannot specify the property in dataset_example.yaml in conf/data_module to get the base model. Even if I specify the property in the yaml file and the training of base model is done, I won't be able to generate new structures with a property condition from the base model. Is that correct?
Therefore, fine-tuning the base model must be done to generate new structures with a property condition. Is that correct?

[danielzuegner]

Hi @Mofahdi,

The base model did not see any property data and therefore cannot be used for conditional generation. Instead of fine-tuning the base model, however, you can also train a conditional model from scratch.

For the band gap example, add the +lightning_module/diffusion_module/model/property_embeddings@lightning_module.diffusion_module.model.property_embeddings.dft_band_gap=dft_band_gap data_module.properties=["dft_band_gap"] overrides to your training command. For instance, this works for me:

mattergen-train data_module=alex_mp_20 ~trainer.logger trainer.accumulate_grad_batches=4 +lightning_module/diffusion_module/model/property_embeddings@lightning_module.diffusion_module.model.property_embeddings.dft_band_gap=dft_band_gap data_module.properties=["dft_band_gap"]

Does this address your issue?

Best,

Daniel