keyky / mobilenet-mxnet Goto Github PK

@KeyKy Would you mind update the fast depthwise operator. or could you describe the key for implement fast operator. Thank you.

Dear writer, ths for your works,
I have a bug when use your model to get a my own fine-tune model,

mxnet.base.MXNetError: Failed loading Op conv2_1_dw of type ChannelwiseConvolution: [19:46:46] D:\Program Files (x86)\Jenkins\workspace\mxnet\mxnet\nnvm\src\core\op.cc:55: Check failed: op != nullptr Operator ChannelwiseConvolution is not registered

So, what's the reason for this, ChannelwiseConvolution is not found?

    conv2_1_sep_bn = mx.symbol.BatchNorm(name='conv2_1_sep_bn', data=conv2_1_sep , use_global_stats=False, fix_gamma=False, eps=0.000100)
    conv2_1_sep_scale = conv2_1_sep_bn
    relu2_1_sep = mx.symbol.Activation(name='relu2_1_sep', data=conv2_1_sep_scale , act_type='relu')

Just curious -- what's the benefit of including the line conv2_1_sep_scale = conv2_1_sep_bn?

In other words, would it be equivalent to have the following?

    conv2_1_sep_bn = mx.symbol.BatchNorm(name='conv2_1_sep_bn', data=conv2_1_sep , use_global_stats=False, fix_gamma=False, eps=0.000100)
    relu2_1_sep = mx.symbol.Activation(name='relu2_1_sep', data=conv2_1_sep_bn , act_type='relu')

I am interested in this implement of the article, but I wonder why you not push you code to
https://github.com/dmlc/mxnet?
and can this modle work? I tried, but failed, I can not compile the whole prject now.

hello, I execute the score.py as below:
1 clone the whole project of yours mobilenet-mxnet
2 cd into mxnet
3 then I execute make command: make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1
4 lastly I execute python score.py --prefix=mobilenet --epoch=0 --data-val=/data/ILSVRC2012_img_test.rec

In this step it went wrong:
(mobilenet) [hanqing@localhost mobilenet-mxnet]$ python score.py --prefix=mobilenet --epoch=0 --data-val=/data/ILSVRC2012_img_test.rec
[21:56:23] src/io/iter_image_recordio_2.cc:135: ImageRecordIOParser2: /data/ILSVRC2012_img_test.rec, use 4 threads for decoding..
[21:56:24] src/nnvm/legacy_json_util.cc:190: Loading symbol saved by previous version v0.9.5. Attempting to upgrade...
[21:56:24] src/nnvm/legacy_json_util.cc:198: Symbol successfully upgraded!
[21:56:25] src/operator/././cudnn_algoreg-inl.h:65: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
[22:01:59] include/dmlc/././logging.h:304: [22:01:59] src/recordio.cc:125: Check failed: pbegin_ <= pend_ Invalid RecordIO Format

Stack trace returned 5 entries:
[bt] (0) /home/hanqing/anaconda2/envs/mobilenet/lib/python2.7/site-packages/mxnet-0.10.1-py2.7.egg/mxnet/libmxnet.so(_ZN4dmlc19RecordIOChunkReader10NextRecordEPNS_10InputSplit4BlobE+0xd86) [0x7f3747c89406]
[bt] (1) /home/hanqing/anaconda2/envs/mobilenet/lib/python2.7/site-packages/mxnet-0.10.1-py2.7.egg/mxnet/libmxnet.so(+0x10366ed) [0x7f37478986ed]
[bt] (2) /lib64/libgomp.so.1(+0x16435) [0x7f372e6e9435]
[bt] (3) /lib64/libpthread.so.0(+0x7dc5) [0x7f375a691dc5]
[bt] (4) /lib64/libc.so.6(clone+0x6d) [0x7f3759cb776d]

terminate called after throwing an instance of 'dmlc::Error'
what(): [22:01:59] src/recordio.cc:125: Check failed: pbegin_ <= pend_ Invalid RecordIO Format

Stack trace returned 5 entries:
[bt] (0) /home/hanqing/anaconda2/envs/mobilenet/lib/python2.7/site-packages/mxnet-0.10.1-py2.7.egg/mxnet/libmxnet.so(_ZN4dmlc19RecordIOChunkReader10NextRecordEPNS_10InputSplit4BlobE+0xd86) [0x7f3747c89406]
[bt] (1) /home/hanqing/anaconda2/envs/mobilenet/lib/python2.7/site-packages/mxnet-0.10.1-py2.7.egg/mxnet/libmxnet.so(+0x10366ed) [0x7f37478986ed]
[bt] (2) /lib64/libgomp.so.1(+0x16435) [0x7f372e6e9435]
[bt] (3) /lib64/libpthread.so.0(+0x7dc5) [0x7f375a691dc5]
[bt] (4) /lib64/libc.so.6(clone+0x6d) [0x7f3759cb776d]

Aborted (core dumped)

Is there anything I do wrong?

I use mxnet-mobilenet for train, but for my GPU 1080Ti, the input is 224*224,my training speed is about 50 sample/second,so is something wrong for my training, I can't find the reason, looking forward your reply

I did not installed the channelwiseconvolution mxnet in your reference.
I convert the model to use convolution group in dmlc/mxnet .
With this I can't get the result accuracy ..

I can't find any paper for the difference.

Can you help me?
Thanks a lot!

I'm a little confused about README.md.

You wrote: This model is converted from MobileNet-Caffe. However, I found the operator your implemented channelwise_convolution.cc,channelwise_convolution.cu in mxnet@c492b5b.
Besides, does this channelwise equal to depth-wise convolution in original paper?

The content below is from orginal paper:

Unstructured sparse matrix operations are not typically faster than dense matrix operations until a very high level of sparsity. 

Our model structure puts nearly all of the computation into dense 1×1 convolutions. This can be implemented with highly optimized general matrix multiply (GEMM) functions. Often convolutions are implemented by a GEMM but require an initial reordering in memory called im2col in order to map it to a GEMM. For instance, this approach is used in the popular Caffe package [15].

1×1 convolutions do not require this reordering in memory and can be implemented directly with GEMM which is one of the most optimized numerical linear algebra algorithms.
MobileNet spends 95% of it’s computation time in 1 × 1 convolutions which also has 75% of the parameters.

So I wann ask, does this repo accomplish with highly optimized GEMM?
Besides, I wann ask: does this need im2col during mobilenet? before normal and depth-wise conv
Thanks!

I noticed that you have used ChannelwiseConvolution in mobilenet-symbol.json file and mobilenet-faster.py. I am using mxnet of version 0.11.1, do i have to use ChannelwiseConvolution with depthwiseConvolution instead? Because your update say that mxnet 0.11.1 support depthwiseConvolution.
Thanks for your reply!