fxmeng/Pruning-Filter-in-Filter

Did the strips really deleted or simply set them to zeros?

Closed this issue · 3 comments

Gaondong commented

Thanks for your works！ But how to prune by strips of a (C*K*K) kernel ?

fxmeng commented

We first set them to zero while training, after that, we pruned all zero stripes, and re-implied the model by stripe-wise convolution layer. (

Pruning-Filter-in-Filter/models/stripe.py

Line 47 in 3207012

def _break(self, threshold):
)

Gaondong commented

We first set them to zero while training, after that, we pruned all zero stripes, and re-implied the model by stripe-wise convolution layer. (

Pruning-Filter-in-Filter/models/stripe.py

Line 47 in 3207012

def _break(self, threshold):

)
I saw the stripe-wise in your paper.
Did the final weights could frozen to .pb format for deploy? The kernel weights which shapes is H*W*Cin*Cout, could not be saved.

fxmeng commented

We change filters from NCKK to (NKK)C11, and can prune each one of the NKK stripes.
However, accelerated deployment with CUDA doesn't provide.
You have to complete it by yourself under models/readme.md's instruction.