yuanyuanli85 / caffemodelcompression Goto Github PK

I am interesting in model compression and acceleration. In the paper(Deep Compression), the authors showed 3x speedup on CPU and 3.5x on GPU. But in here, I can't accelerate my AlexNet but only do compression and decompression. How to implement acceleration in the paper? Have any CNN accelerator in Caffe?
Thinks

Thank you very much for sharing you code . But when I test, the inferece time for lenet or pvanet didn`t speed up. But accuracy is very lost.

Hi @yuanyuanli85
thank you for your work and share.
It looks very promising wrt to model size after compression. I would like to use compressed caffemodel, do you know how can i use it if I have hardware which is capable of low-bit calculation

-anand

Hello

I'm so interested for your project.
I have tried your code for VGG16 net, but why compress output become npz ?
Can i use this npz output to be pre trained model for my caffe ?

thanks

I got the following error while trying to build.

gcc: error: unrecognized command line option ‘-mavx2’
/usr/bin/ld: weights_compress.o: Relocations in generic ELF (EM: 62)
/usr/bin/ld: weights_compress.o: Relocations in generic ELF (EM: 62)
/usr/bin/ld: weights_compress.o: Relocations in generic ELF (EM: 62)
/usr/bin/ld: weights_compress.o: Relocations in generic ELF (EM: 62)
weights_compress.o: error adding symbols: File in wrong format

I want to decompress the model with c++, I see that this project is based on c++. Can you provide the c++ interface? Thanks.

Thanks for the code, I'm hoping it will be useful for my work.

It would be very helpful if you could provide some instructions on how to use this for the general use case, where someone wants to compress a previously trained model, fine-tune the compressed model, iterate as needed, and then use it for testing. Is this possible with the code that you provided, assuming caffe is already installed and working?

Thanks again.

你好，非常感谢您提供代码。虽然我在进行物体检测研究时，用您的代码将我的模型压缩了十倍。从100M压到了11M，测试的时候也能被caffe使用，，但是速度明显下降，测试一张图片要3.5秒，这是不可接受的，请问，这是什么原因造成的？有解决办法吗

Hi, Thanks for you codes. What is the following sentence function to achieve?
caffe_model_decompress (netmodel, "alexnet_xx.caffemodel", "alexnetzip.npz")
Why is the output of alexnet_xx.caffemodel the size is not the same of the not compressed before? How to fine tune the data set test accuracy with the compressed model？I will be very grateful if anyone can help me, Thanks!

There's an index out of bounds error when trying to compress layers which do not include bias parameters. I can provide a solution via PR until next week.

I want to store and quantify the sparse matrix that has been pruned for AlexNet, and there is no problem when I run./build.sh.However, when I ran python caffemodel_compress. Py, the following error occurred. I tried all kinds of methods but couldn't find the problem

@yuanyuanli85 @marifnst @gavenswang

Hi,
I have a trouble about network compression. for alexnet, we can compressed it, but how can we use the compressed model? If we decompress it to use, then we also get a big file and the speed was not speed up. my question is how to use the compressed model directly?

Thank you for sharing the code. There is a question which still puzzled me. It will be helpful if you can explain the meaning of the parameter "newlabels" .

I am compress the https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
and decompress it, then use this decompressed model file for testing.
I have tested several images, everything works well except this image:
caffe/examples/images/cat.jpg (https://github.com/BVLC/caffe/blob/master/examples/images/cat.jpg)

give the result of:
output label: n02119022 red fox, Vulpes vulpes

thanks for your code ,but when i run the code ,i got this error:
Traceback (most recent call last):
File "caffemodel_compress.py", line 96, in
caffe_model_compress(netmodel, netweights, output, 6, 2)
File "caffemodel_compress.py", line 20, in caffe_model_compress
for item in net.params.items():
File "/home/wl/caffe/python/caffe/pycaffe.py", line 68, in _Net_params
if len(lr.blobs) > 0])
OverflowError: long int too large to convert to int
help me please!

Sorry, I do not know how to use .npz file as caffemodel to test. Could you please tell me?

Thank you for your code!
when I run the code ,i got an error: illegal instruction (core dump)
I0405 19:50:28.866523 7248 upgrade_proto.cpp:61] Successfully upgraded file specified using deprecated V1LayerParameter
I0405 19:50:28.964052 7248 net.cpp:744] Ignoring source layer loss
compressing layer conv1
test1 here
test2 here
test3 here
illegal instruction (core dump).
i set flags(test* here) to check ,
codebook = np.empty((2**nbit),dtype=np.float32)
print "test3 here"
#t_start = time.time()
wqtz.compress_layer_weights(newlabel, codebook, weights_vec, vec_length, nbit)
#t_stop = time.time()
#kmeans_time = kmeans_time + t_stop - t_start
print "test4"
it seems that error comes from compress_layer_weights().
how can I solve it.