reproducebyyq / channel_pruning_yq Goto Github PK

When I test it ,after lasso regression ,I do this:
count = sum(idxs)
idxs[0:count]=True
idxs[count:]=False
the number of filters is the same , but the first count filters to preserve. the top5 and top1 is no descend ,why?

thanks for your sharing, is simpler to understand than oringal code. then,the question is as follows:
{
"conv1_2": 24,
"conv2_1": 48,
"conv2_2": 48,
"conv3_1": 64,
"conv3_2": 128,
"conv3_3": 160,
"conv4_1": 192,
"conv4_2": 192,
"conv4_3": 256,
"conv5_1": 320,
"conv5_2": 320,
"conv5_3": 320
}

How to calculate and get this number?

Hi,thanks for sharing, but I get the following error when I run your code:
Traceback (most recent call last):
File "./low_rank_and_channel_pruning.py", line 12, in
from util import * # my function
ImportError: No module named 'util'
Can you share this file you wrote?thanks

Hi, thank you for your code to help me understand the author's long code, but I have some question about your code.
First，why the "nPicsPerBatch" in your code is k_h*k_w? And "N" in your code is "N = nBatch * nPointsPerLayer" ,could you please explain it?
Second, is there any mistake in the code [ https://github.com/ReProduceByYQ/Channel_pruning_yq/blob/54bb57b96b884d6ae4b2631c7e92150322cabb3e/util.py#L198] ? When I run your code ,there is an error , since the randx=randy=10,then the point_feat should be (1,48, 10, 10). After reshaping, the size should be(480,10). Is that right? I am hoping for your reply. Thank you so much!

剪枝resnet实现了没有，看了下原文实现好像将caffe底层也改了，我想不改底层，能否给点建议。

Hi Jason Yue,

Your code helps me A LOT (^_^) during understanding He Yihui's channel pruning paper. My question is, have you try fully prune VGG16 (from conv1_1 to conv4_3) using your implementation? Did you manage to achieve only 1.7% Top5 accuracy lost under 5X as the paper reported?

I also re-implemented He Yihui's paper with only channel pruning, but when I prune VGG16 for 5X (pruning each layer's channel number to the exact same number of He Yihui's 5X caffemodel prototxt), and the Top5 accruacy loss is around 5%, and the paper repored 1.7% only.

Did you also experience this issue? Thanks.

压缩的时候使用的是验证集，测试的时候还是验证集，这样合理吗？