Pruning Rate Control

[yeeunleee]

Hello, thank you for your great work!

In Table 2, pruning rates are fixed at 30%, 50%, and 70%.
However, since SeTa and InfoBatch dynamically change the number of samples per epoch.
(e.g., varying cluster sizes in SeTa)
It seems difficult to precisely control the pruning rate.

Could you explain how you matched these target pruning ratios?

[mrazhou]

Thank you for your interest in our work.

It is indeed challenging to precisely control the final, overall pruning rate for dynamic methods like SeTa and InfoBatch. To achieve those target ratios, we primarily relied on hyperparameter grid search and selected the configurations that yielded final rates closest to our targets (for example, accepting 29% or 31% as an approximation for a 30% target).

Because experiments on CIFAR-10/100 run quite fast, this parameter sweeping approach is highly feasible.

If you are setting this up for SeTa, here are a few practical tips:

• Cluster Size: You don't need to tune this much; 5 or 10 works perfectly fine.
• Targeting Low/Medium Rates: Set the window to 0.9 and adjust the --ratio parameter starting from 0.0.
• Targeting High Rates (e.g., 70%): We recommend lowering the window parameter to 0.6.

If your goal is simply to extract results at various pruning rates for baseline comparisons, I recommend checking out Figure 3(b) in our paper, which illustrates these trends.

On a related note, we recently published a new method called RePB (ICLR2026, paper, code). It requires fewer hyperparameters to tune compared to SeTa and is generally more stable. We'd love for you to give it a try!

Please feel free to follow up here or reach out via email if you have any further questions.