Questions on DSCF pretraining code and COREL implementation

[audreyeternal]

Thank you for providing the foundation model DSCF for spectrometry.

I have a few questions regarding the code and pretraining:

1. Is there pretraining and inference code available? The loss function described in Eq. (1) of the paper doesn’t seem to be implemented in the current code.

2. The paper states:

“COREL involves sampling and synthesizing spectral data from the spectral dictionary and using spectral component filtering methods to separate the spectral components present in complex systems.”

Just to confirm: does the objective of the pre-training (COREL) is to take in a mixture spectrum (possibly with a random mask) and aim to separate components?
If MultiDec_1d_model is the pretraining model, how should the outplanes parameter be set? The default seems to be 1, which reduces it to a standard MAE task.

3. From the code, it looks like spectral component abundances in mixtures are randomly sampled. Could this affect performance due to mismatch with real-world abundance distributions?

Thanks in advance!

[streamflowmaster]

Thank you for providing the foundation model DSCF for spectrometry.

I have a few questions regarding the code and pretraining:

1. Is there pretraining and inference code available? The loss function described in Eq. (1) of the paper doesn’t seem to be implemented in the current code.
2. The paper states:

“COREL involves sampling and synthesizing spectral data from the spectral dictionary and using spectral component filtering methods to separate the spectral components present in complex systems.”

Just to confirm: does the objective of the pre-training (COREL) is to take in a mixture spectrum (possibly with a random mask) and aim to separate components? If MultiDec_1d_model is the pretraining model, how should the outplanes parameter be set? The default seems to be 1, which reduces it to a standard MAE task.

3. From the code, it looks like spectral component abundances in mixtures are randomly sampled. Could this affect performance due to mismatch with real-world abundance distributions?

Thanks in advance!

Thanks for your feedback!

The inference code and fine-tuning code are available in the repository. The loss function described in Equation (1) of the paper is implemented in the file customized_task/finetune_engine.py as the functions pairwise_cos_sim and cal_loss. The logarithmic pairwise_cos_sim loss should be reactivated after the initial training steps to ensure stable training.

Your claim is generally correct. COREL models complex mixtures by filtering components. The outplanes parameter can be set via DSCFConfig as a positive integer, with a default value of 1, which simplifies the model to use modified mask modeling methods.

Yes, you are correct; component abundances are sampled in two steps: random components and random abundances. This is an excellent question, and I have indeed considered its potential influence. Experimental results show that sufficient randomness can approximately cover real-world abundance distributions, although it requires more computational resources. The concept of real-world abundance distributions exists but is somewhat idealized, making it impractical to fully characterize during pretraining. These distributions should be carefully and rigorously considered when fine-tuning for your specific downstream tasks in the future.