Questions on Fine-Tuning Parameter Percentage Calculation

[mrclovvn]

Dear authors,

Thank you for your great work and the publicly available code! I noticed in Table 3 of the paper that the fine-tuning parameter percentage for PYRA in ViT-B/16 is listed as 0.35%. Could you please clarify how this percentage is calculated?

When I try to replicate the results using the experiments/LoRA/ViT-B_prompt_lora_8.yaml configuration from the public code, I get the following parameter information:

total training parameters: 399652 adapter 0 LoRA 294912 prompt 0 prefix 0 PYRA 27840 head 76900
total parameters in model: 86198308

However, when I calculate the percentage as (LoRA + PYRA) / total, I get (294912 + 27840) / 86198308 = 0.3744%. Alternatively, when I calculate it as (LoRA + PYRA) / (total - head), I get (294912 + 27840) / (86198308 - 76900) = 0.3748%, which is different from the value of 0.35% in Table 3.

Similarly, when I use the experiments/LoRA/ViT-L_prompt_lora_12.yaml configuration to fine-tune ViT-L, the parameter information is as follows:

total training parameters: 1356068 adapter 0 LoRA 1179648 prompt 0 prefix 0 PYRA 73920 head 102500
total parameters in model: 304657700

The percentage I calculate is (1179648 + 73920) / 304657700 = 0.4115% and (1179648 + 73920) / (304657700 - 102500) = 0.4116%, which again does not match the value of 0.40% in Table 3.

Could you please explain how the fine-tuning parameter percentage is computed in the paper? Am I misunderstanding the calculation process?

Thank you for your time and assistance!

[Bostoncake]

Hello! Sorry for the misunderstanding between our codebase and the article. When we were working on the draft, we have adopted a version of PYRA without the LayerNorm, and the numbers in the paper corresponds to the LayerNorm-free version.

You are correct and that was a good catch. We will update the draft on arXiv accordingly.

[mrclovvn]

Thank you for your prompt response and clarification! I look forward to the revised version on arXiv.