Filter prune algo implementation #1655
Conversation
chicm-ms
requested review from
liuzhe-lz,
LinquanLiu,
QuanluZhang and
xuehui1991
and removed request for
LinquanLiu
| @@ -0,0 +1,130 @@ | |||
| from nni.compression.tensorflow import FPGMPruner | |||
There was a problem hiding this comment.
is it easy to reproduce one of the experiments in the paper?
There was a problem hiding this comment.
The key metrics of the pruning are reduced FLOPs and accuracy change, to reproduce the reduced FLOPs, we need to use the pruned compact model, which is not implemented yet.
There was a problem hiding this comment.
I think we need to offer some test metrics to make sure the implementation is correct.
There was a problem hiding this comment.
Sure, I have added verification code in fpgm_torch_mnist.py example to verify the pruned conv kernel weight sparsity. By checking the sparsity and loss, we can verify:
- the configured layers are pruned.
- the pruned model has similar loss with origin model
But this code still can not verify the implementation is same as the paper. I am considering to add some kind of verification in UT.
|
please update doc accordingly |
| } | ||
|
|
||
| default_layers = ['Dense', 'Conv2D'] |
There was a problem hiding this comment.
what is the difference between supported_layers and default_layers?
There was a problem hiding this comment.
Thanks, updated, please check again.
Filter prune algo implementation (microsoft#1655)
Filter prune algo implementation (microsoft#1655)
This reverts commit 262fabf.
Implement filter pruning algorithms: Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration (https://arxiv.org/abs/1811.00250)