Question about parameter count

[brshook2023]

In your paper, you report that CBraMod has 4M parameters. However, when the model is loaded and finetuned, the number of trainable parameters is much higher, depending on the dataset. For example,

• Shu: 25.5M
• Physio: 46M
• BCICIV2a: 19.1M
• Stress: 25.1M
• SEEDV: 15M

For reference, I use total_params = sum(p.numel() for p in model.parameters() if p.requires_grad) to get the number of trainable parameters. Could you explain the discrepancy here? Thank you!!

[g12bftd]

Hey, I also noticed a similar pattern. From what I understand, the CBraMod encoder has about 4M params. However, if you look at the classifier/regression head, it adds significantly more parameters. You can double check this by running the total_params calculation on the classifier alone.