younghwanoh/impl-pruning-caffemodel

Simple pruning example using Caffe

impl-pruning-caffemodel's Issues

Can't prune more than 3 layers.

I changed the source code to include more layers in the ratio array, but only 3 layers are being altered. How do you prune more layers?

ValueError

Hello, when I am running proto.py bvlc_alexnet.caffemodel output_pruned.caffemodel encounter the following mistakes, what is the reason and how to solver the problem, thanks
dl@dl:~/home/impl-pruning-caffemodel-master$ python proto.py bvlc_alexnet.caffemodel output_pruned.caffemodel
layer name: fc8
width: 4096
height: 1000
pruning off 75.0 % of this layer

Traceback (most recent call last):
File "proto.py", line 51, in
i.blobs[0].data.extend(temp)
File "/usr/lib/python2.7/dist-packages/google/protobuf/internal/containers.py", line 125, in extend
if not elem_seq:
ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

using pycaffe.py to prune the alexnet model, but no performance improved

I downloaded the alexnet from: https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet

Using the pycaffe.py to prune this net:
python pycaffe.py deploy.prototxt bvlc_alexnet.caffemodel alexnet_pruned.caffemodel

Then I use the alexnet_pruned.caffemodel for predicting the image using caffe python:

t_start = time.time()                                                                                                                                                       
output = net.forward()                                                                                                                                                      
t_stop = time.time()                                                                                                                                                        
duration = t_stop - t_start  #performance

I have tested in both CPU and GPU mode(caffe.set_mode_cpu() & caffe.set_mode_gpu() ), the performance seems the same.

I have some questions that need to be answered.

Hi, I use freeze.py on the VGG model, but the size of output.caffemodel is the same as the original size. Is this the result right?