cornell-zhang / dnn-gating Goto Github PK

Hello,

Great work! Thanks for sharing the code, I would like to inquire about how the sparsity is reported [https://github.com/cornell-zhang/dnn-gating/blob/master/utils/cg_utils.py#L176]. This calculates sparsity in feature maps per pixel, but as far as I know, this isn't how the filter pruning methods report FLOPs reduction. I was wondering if there is another FLOPs reduction report that calculates based on skipping the whole filter/slice or not to be comparable with filter pruning methods instead of sparsity measure?

Also, in your paper table 4, you calculated fps with batch size 32. Would you mind sharing code to achieve actual speedup?

Thanks!

Did you implement two separate BN for base path (Xp) and conditional path (Xr) as figure 4 mentioned in your paper?
I only find one BN here.

In your implementation of QuantizedConv2d, I found your pipeline is: quantize -> dequantize -> conv, why not use quantize -> conv -> dequantize?

in my opinion, the pipeline (quantize -> conv -> dequantize) will be faster, because the conv can be boosted by low-bit calculation.

Hope to get your reply.
Thanks

Did you implement threshold tc mentioned in formula (4) in this repo?
I think the line below only implements Δ

According to the code here, the implementation of CGConv2d does not skip the computation of Y_r, which is proposed in your paper.

In fact, this implementation computes Y and Y_p separately, and produces the result by Y*d+Y_p*(1-d). This implementation increases the computation, right? So how do you actually skip and save the computation?

Thanks!