Hi,
I Pruned a Tensorflow based MobileneV2 classification model using structured pruning (BYOM Approach) with different values for the min.no.of filters per layer - specifically 8 and 16. After retraining both pruned models , I noticed no difference in performance (both accuracy and size seems nearly identical).
Why this happen, and what is the exact role of this parameter in pruning ?
Hi @Aryasree_S,
I’m not an expert here. But since you’re specifying ‘minimum’ number of filters per layer I’d expect that the actual number of filters is greater than 16 - Setting the minimum to 8 vs 16 isn’t changing the resultant model because of this.
Best,
Sophie
Hi @sophwats ,
What would be the best pruning threshold to apply? Additionally is it advisable to try pruning with min.no.of filters greater than 16?
Best,
Aryasree S
Hi @Aryasree_S,
Each of the thresholds and parameters can impact the size of the model and accuracy, and ‘optimal’ values depend on your specific model. You need to consider these tradeoffs of each when selecting a value.
Best,
Sophie
1 Like
The min.no.of filters parameter acts as a lower bound to prevent excessive pruning in sensitive layers. If pruning fails to remove meaningful filters, the identical post-pruning sizes suggest no difference in performance.
1 Like
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.