add symbol.SwapAxis operator, just can do Forward().

[starimpact]

this is a early version, can run now, but haven't been tested.
This is just a look-version, means let you have a look and give me some ideas.

[starimpact]

I am sorry for did not see your comments.

[starimpact]

thanks for your suggestions.

[starimpact]

I have finished the new code based on your suggestions.
And the Backward is also completed.
Please check the code for me.Give me more suggestions.
Thank you very much.:smile:

[tqchen]

Please change this back to default, as most users don't have cuda

[tqchen]

Thanks for the contribution. I have made a few comments on the code, in general

• Code Style, we use Google C++ code style, you can reproduce the linter check locally using

make lint

• Try to use standard functions when possible
  • This include std::swap and std::accumulate
• DeclareBackwardDependency, as Backward only depend on out_grad
  • c.f. http://mxnet.readthedocs.org/en/latest/developer-guide/operator.html
• Please also write a testcase in https://github.com/dmlc/mxnet/blob/master/tests/python/unittest/test_operator.py

[starimpact]

Thanks very much for your suggestions.

[starimpact]

function is ready to go. please check it.
maybe there is a little code style problem.
😆 😄 😸 😈

[starimpact]

I come across a problem:
SwapAxis(..., dim1=2, dim2=3)can work well on both cpu and gpu.
SwapAxis(..., dim1=3, dim2=2)only work well on cpu, but fail on gpu with the following error:

INFO:root:Start training with [gpu(0)]
[13:46:12] ./dmlc-core/include/dmlc/logging.h:208: [13:46:12] ./mshadow/mshadow/./tensor_blob.h:530: Check failed: (this->shape_.Size()) == (shape.Size()) TBlob.get_with_shape: new and old shape do not match total elements
[13:46:12] ./dmlc-core/include/dmlc/logging.h:208: [13:46:12] src/engine/./threaded_engine.h:295: [13:46:12] ./mshadow/mshadow/./tensor_blob.h:530: Check failed: (this->shape_.Size()) == (shape.Size()) TBlob.get_with_shape: new and old shape do not match total elements
terminate called after throwing an instance of 'dmlc::Error'
  what():  [13:46:12] src/engine/./threaded_engine.h:295: [13:46:12] ./mshadow/mshadow/./tensor_blob.h:530: Check failed: (this->shape_.Size()) == (shape.Size()) TBlob.get_with_shape: new and old shape do not match total elements
Aborted (core dumped)

[starimpact]

std::accumulate can not be recognized by nvcc.

[tqchen]

Oh, yap. Maybe create our own version of prod function is easier. Then I think it is OK. Thanks

[tqchen]

For the error, as it indicates, it means the shape's size do not match the size of TBlob. Likely due to shape initialization error, either InferShape or Shape2Five

[tqchen]

The platform dependency issue was likely due to some uninitialized memory(variable) that causes uncertainties, but just my guess

[tqchen]

resolve all the comments by cpplint, there are detailed messages

[tqchen]

For output, we need to do

exec_c.forward()
out = exec.output[0].asnumpy()
print out

The error was due to we did not call asnumpy to wait for the result, and the system start to shutdown before the computation starts.

[starimpact]

I can see nothing comments in the swapaxis files by cpplint? where are they?

[starimpact]

there is something error...

=====69/70 cpp-header files passed check=====
src/operator/swapaxis-inl.h: 34 Errors of 4 Categories map={'whitespace': 29, 'runtime': 2, 'readability': 2, 'build': 1}
=====57/58 cpp-soruce files passed check=====
src/operator/swapaxis.cc: 2 Errors of 1 Categories map={'readability': 2}
=====40/40 python files passed check=====
2 files failed lint
make: *** [lint] Error 1

[tqchen]

Hmm the error message occurs before these when it scan through files. You
may need to scroll up a bit to see it
On Sat, Nov 7, 2015 at 10:46 PM Zhang Ming [email protected] wrote:

there is something error...

=====69/70 cpp-header files passed check=====
src/operator/swapaxis-inl.h: 34 Errors of 4 Categories map={'whitespace': 29, 'runtime': 2, 'readability': 2, 'build': 1}
=====57/58 cpp-soruce files passed check=====
src/operator/swapaxis.cc: 2 Errors of 1 Categories map={'readability': 2}
=====40/40 python files passed check=====
2 files failed lint
make: *** [lint] Error 1

—
Reply to this email directly or view it on GitHub
#502 (comment).

[starimpact]

cpu is work right now, but not for gpu:

def test2():
    data_in = mx.symbol.Variable('data')
    conv = mx.symbol.Convolution(data=data_in, kernel=(3, 3), num_filter=16)
    datatmp = np.ones((1, 1, 32, 64))
    mxdata = mx.nd.array(datatmp)
    weightmp = np.ones((16, 1, 3, 3))
    mxweight = mx.nd.array(weightmp)
    biastmp = np.zeros(16)
    mxbias = mx.nd.array(biastmp)
    exe_c = conv.bind(ctx=mx.gpu(0), args=[mxdata, mxweight, mxbias])
    exe_c.forward()
    out = exe_c.outputs[0].asnumpy()
    print out

test2()

Error

[14:50:16] ./dmlc-core/include/dmlc/logging.h:208: [14:50:16] ./mshadow/mshadow/./tensor_blob.h:508: Check failed: Device::kDevMask == dev_mask_ && DataType<DType>::kFlag == type_flag_ TBlob.get: device type do not match specified type
[14:50:16] ./dmlc-core/include/dmlc/logging.h:208: [14:50:16] src/engine/./threaded_engine.h:295: [14:50:16] ./mshadow/mshadow/./tensor_blob.h:508: Check failed: Device::kDevMask == dev_mask_ && DataType<DType>::kFlag == type_flag_ TBlob.get: device type do not match specified type
terminate called after throwing an instance of 'dmlc::Error'
  what():  [14:50:16] src/engine/./threaded_engine.h:295: [14:50:16] ./mshadow/mshadow/./tensor_blob.h:508: Check failed: Device::kDevMask == dev_mask_ && DataType<DType>::kFlag == type_flag_ TBlob.get: device type do not match specified type
Aborted (core dumped)

[starimpact]

make lint is all passed!

[tqchen]

CHECK_EQ(input_shape->size(), 1);

[tqchen]

Thanks for the good job! I have last few comments, please address them and rebase to resolve the conflict to current master here.

http://mxnet.readthedocs.org/en/latest/contribute.html#how-to-resolve-conflict-with-master

The conflict might have something to do with commits you have in mshadow, if that is the case, the best way might be reset mshadow's version, or keep a copy of your files, and do a clean fork

[starimpact]

Done, please have a check!

[starimpact]

oh, sorry, have not updated your newest comments.

[starimpact]

I think you forgot I can not push to your repository.

[starimpact]

Do I have another way to add your newest commits without do second time fork?

[tqchen]

You can do a rebase http://mxnet.readthedocs.org/en/latest/contribute.html#how-to-resolve-conflict-with-master

[starimpact]

something is wrong when I do rebase your mxnet.

[starimpact]

I do fetch your mxnet, but my repository is changed to my older version one when I do rebase.
really strange.

[starimpact]

I have pushed the newest code to my forked mxnet, can you check it?

[starimpact]

I need a little time to figure out the strange problem of rebase.

[starimpact]

Can I do merge?

[tqchen]

The general instruction is here http://mxnet.readthedocs.org/en/latest/contribute.html#how-to-resolve-conflict-with-master

If you find files with conflicts, edit the files and merge the conflict, and do a git add as indicated in the instruction

[starimpact]

the situation is: I have commit a lot in my local repository, when I rebase to your newest master, my current work directory content is updated to the first commit. What can I do?

[starimpact]

OK, likely I found the way....

[starimpact]

ok ,now, please check.

[tqchen]

closed due to #519