gluon improvement #8152

szha · 2017-10-05T03:45:08Z

changes include:

use generator expr for arguments when calling cached op
merge two separate for loops for assembling parameters into one so that the check for parameter membership in cargs is cached.
fill shapes after parameter data is set, to be reflected in block's __repr__.

piiswrong · 2017-10-05T04:24:41Z

Construction is only called once

piiswrong · 2017-10-05T04:24:53Z

python/mxnet/gluon/block.py

-        assert fmt == self._in_format, "Invalid input format"
-        for i, j in self._in_idx:
-            cargs[i] = args[j]
+            cargs = tuple(args[self._in_idx[i]] if self._in_idx[i] != -1 else v for i, v


This doesn't look right

This is a simple combination of the list comprehension and the for loop. What seems to be the problem?

All tests passed.

Could you double check and see if still doesn't look right? If so, what looks like the error? And what other tests should be added to catch the error?

piiswrong · 2017-10-05T18:27:05Z

What's the benefit of this change? This part is complicated and hard to debug, so unless there is visible performance impact I think there is no reason to change it

mli · 2017-10-05T20:14:36Z

Agree that using tuple and iterator can reduce the memory. But Eric's point is valid, this PR puts too many things in a single expression (i.e. cargs = ...), which makes the codes harder to read.

maybe we can write a generator for cargs, which can further reduce memory, and break single line into multiple lines to improve readability.

also, is there any benchmark showing that this kinds of modification can improve things in a notiable way?

szha · 2017-10-05T23:00:22Z

Thanks for the comments, I worked on this a bit more and simplified the assembly of calling arguments using one loop. I renamed the field for readability. And here are the timing results:

workload:

import mxnet as mx
from mxnet import gluon
class example(gluon.HybridBlock):
    def __init__(self, num_args, num_params):
        super(example, self).__init__()
        self._num_args = num_args
        self._num_params = num_params
        for i in range(num_params):
            setattr(self, 'param%d'%i, self.params.get('param%d'%i, shape=(1,), init='ones'))
    def hybrid_forward(self, F, *args, **kwargs):
        return F.concat(*(a for a in (args+tuple(kwargs.values()))), dim=0)

num_args, num_params = 2, 50
net = example(num_args, num_params)
net.initialize()
net.hybridize()

And run:

%timeit -n 100 net(*(mx.nd.ones((1,)) for _ in range(num_args))).wait_to_read()

Results:

num_args, num_params = 100, 200
# before
100 loops, best of 3: 5.52 ms per loop
# after
100 loops, best of 3: 5.37 ms per loop

num_args, num_params = 2, 50
# before
100 loops, best of 3: 333 µs per loop
# after
100 loops, best of 3: 312 µs per loop

Also, I'm caching the result of collect_params now so that checkpointing will be faster.

piiswrong · 2017-10-07T21:50:23Z

Check pointing is not a bottleneck and collecting param is unlikely a big cost during checkpointing.

This is over complicated. Optimization should be driven by profiling. Premature optimization is the source of all evil.

When a layer is changed, simply clear cached op.

szha · 2017-10-09T20:08:42Z

Reverted per request and ready for review.

* Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from #8152 * Add tests from #8152

* Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152

szha · 2017-10-22T05:33:29Z

added shape completion after loading parameters or finishing deferred init in this PR, since the logic for shape completion can benefit from the refactoring of _finish_deferred_init in this PR.

cjolivier01 · 2017-10-22T06:31:49Z

Premature optimization is the root of ALL evil? Wow. :)

szha · 2017-10-22T06:37:11Z

yeah, thou shalt yield when Knuth is quoted.

* CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (#8125) * v0.12 regression: Fix registration of children for Block (#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from #8152 * Add tests from #8152 * Revert "[CMAKE] Fix windows cmake build" (#8311) * Revert "Added my code signing key (#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (#8300) * Update rnn.md (#8320) * fluent methods for missed ops (#8329) * update ps lite (#8327) * Fix unused type warning (#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (#8369) * Allow test to converge (#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (#7988) * [Perl] emulate Python zip() for Perl (#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (#8376) Fix a typo in the example readme.

* CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (apache#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (apache#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (apache#8125) * v0.12 regression: Fix registration of children for Block (apache#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152 * Revert "[CMAKE] Fix windows cmake build" (apache#8311) * Revert "Added my code signing key (apache#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (apache#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (apache#8300) * Update rnn.md (apache#8320) * fluent methods for missed ops (apache#8329) * update ps lite (apache#8327) * Fix unused type warning (apache#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme.

* Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152

* CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (apache#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (apache#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (apache#8125) * v0.12 regression: Fix registration of children for Block (apache#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152 * Revert "[CMAKE] Fix windows cmake build" (apache#8311) * Revert "Added my code signing key (apache#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (apache#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (apache#8300) * Update rnn.md (apache#8320) * fluent methods for missed ops (apache#8329) * update ps lite (apache#8327) * Fix unused type warning (apache#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme.

* Fill optimizations * Optimize IdentityCompute for CPU * lint * Fix unused type warning (#8316) * remove unused variable * CR comments * CR comments * Added _full operator * Trigger build * Trigger build * Add _full to symbolic * Merge conflict resolution fix * lint * Timing output for test_factorization_module when Verbose enabled (#8363) * Timing output for test_factorization_module when Verbose enabled * Trigger build * Trigger build * Trigger build * Misc fixes for sparse distributed training (#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (#8369) * Allow test to converge (#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (#7988) * [Perl] emulate Python zip() for Perl (#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (#8376) Fix a typo in the example readme. * Use omp_get_max_threads() when OMP_NUM_THREADS environment variable is set (#8379) * CPU optimization for ActivationOp (#8296) * CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (#8125) * v0.12 regression: Fix registration of children for Block (#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from #8152 * Add tests from #8152 * Revert "[CMAKE] Fix windows cmake build" (#8311) * Revert "Added my code signing key (#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (#8300) * Update rnn.md (#8320) * fluent methods for missed ops (#8329) * update ps lite (#8327) * Fix unused type warning (#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (#8369) * Allow test to converge (#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (#7988) * [Perl] emulate Python zip() for Perl (#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (#8376) Fix a typo in the example readme. * Fix GPU copy * Remove duplicate * Trigger build

* Memory set/copy speed assertions * Memory set/copy speed assertions * .. * .. * .. * .. * bounce some cache * lint * Timing output for test_factorization_module when Verbose enabled (#8363) * Timing output for test_factorization_module when Verbose enabled * Trigger build * Trigger build * Trigger build * Misc fixes for sparse distributed training (#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (#8369) * Allow test to converge (#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (#7988) * [Perl] emulate Python zip() for Perl (#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (#8376) Fix a typo in the example readme. * Use omp_get_max_threads() when OMP_NUM_THREADS environment variable is set (#8379) * CPU optimization for ActivationOp (#8296) * CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (#8125) * v0.12 regression: Fix registration of children for Block (#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from #8152 * Add tests from #8152 * Revert "[CMAKE] Fix windows cmake build" (#8311) * Revert "Added my code signing key (#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (#8300) * Update rnn.md (#8320) * fluent methods for missed ops (#8329) * update ps lite (#8327) * Fix unused type warning (#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (#8369) * Allow test to converge (#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (#7988) * [Perl] emulate Python zip() for Perl (#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (#8376) Fix a typo in the example readme. * do gtest test * add assert and do higher runs as performance test only (when performance test flag set) * Trigger build * lint * Trigger build * Sparse operator performance improvement (#8412) * sparse rsprsp perf improvements * Clean up * dtype default to source_array.dtype for sparse ndarrays (#8403) * derive default dtype/ctx from input for sparse ndarrays * add gpu tests * fix lint. add doc * remove default_ctx code * bug fix when passing dtype to array() * update doc * remove extra line * also check ctx * fix using default mean pixels (#8352) * fix gluon.data.RecordFileDataset (#8353) * upgrade MKL (#8378) * Lint fix (#8402) * Trigger build

piiswrong · 2017-11-03T04:19:58Z

python/mxnet/gluon/rnn/rnn_layer.py

-        mapping = ('{_input_size} -> {_hidden_size}'.format(**self.__dict__) if self._input_size
-                   else self._hidden_size)
+        shape = self.i2h_weight[0].shape
+        mapping = ('{0} -> {1}'.format(shape[1], shape[0]) if shape[1]


It should show None -> xx when shape is not defined yet

* Fill optimizations * Optimize IdentityCompute for CPU * lint * Fix unused type warning (apache#8316) * remove unused variable * CR comments * CR comments * Added _full operator * Trigger build * Trigger build * Add _full to symbolic * Merge conflict resolution fix * lint * Timing output for test_factorization_module when Verbose enabled (apache#8363) * Timing output for test_factorization_module when Verbose enabled * Trigger build * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme. * Use omp_get_max_threads() when OMP_NUM_THREADS environment variable is set (apache#8379) * CPU optimization for ActivationOp (apache#8296) * CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (apache#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (apache#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (apache#8125) * v0.12 regression: Fix registration of children for Block (apache#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152 * Revert "[CMAKE] Fix windows cmake build" (apache#8311) * Revert "Added my code signing key (apache#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (apache#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (apache#8300) * Update rnn.md (apache#8320) * fluent methods for missed ops (apache#8329) * update ps lite (apache#8327) * Fix unused type warning (apache#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme. * Fix GPU copy * Remove duplicate * Trigger build

* Memory set/copy speed assertions * Memory set/copy speed assertions * .. * .. * .. * .. * bounce some cache * lint * Timing output for test_factorization_module when Verbose enabled (apache#8363) * Timing output for test_factorization_module when Verbose enabled * Trigger build * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme. * Use omp_get_max_threads() when OMP_NUM_THREADS environment variable is set (apache#8379) * CPU optimization for ActivationOp (apache#8296) * CPU optimization for ActivationOp Significant improvement on CPU (several magnitudes of order in some cases, especially on backward pass). Very slight improvement on GPU. OLD MSHADOW APPROACH -------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 18.948 ms, avg: 0.037896 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.658 ms, avg: 0.003316 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 57.973 ms, avg: 0.115946 ms X 500 passes Activation Operator CPU: Timing [Backward] 4.748 ms, avg: 0.009496 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 703.446 ms, avg: 1.40689 ms X 500 passes Activation Operator CPU: Timing [Backward] 56.255 ms, avg: 0.11251 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 2107.77 ms, avg: 4.21554 ms X 500 passes Activation Operator CPU: Timing [Backward] 168.483 ms, avg: 0.336966 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 24122.2 ms, avg: 48.2443 ms X 500 passes Activation Operator CPU: Timing [Backward] 1908.7 ms, avg: 3.8174 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.637 ms, avg: 0.003274 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.665 ms, avg: 0.00333 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.562 ms, avg: 0.003124 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.661 ms, avg: 0.003322 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.635 ms, avg: 0.00327 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.702 ms, avg: 0.003404 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.83 ms, avg: 0.00366 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.041 ms, avg: 0.004082 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.08 ms, avg: 0.00416 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.688 ms, avg: 0.005376 ms X 500 passes NEW MXNET_OP APPROACH --------------------- CPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator CPU: Timing [Forward] 80.748 ms, avg: 0.161496 ms X 500 passes Activation Operator CPU: Timing [Backward] 1.176 ms, avg: 0.002352 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator CPU: Timing [Forward] 7.881 ms, avg: 0.015762 ms X 500 passes Activation Operator CPU: Timing [Backward] 2.181 ms, avg: 0.004362 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator CPU: Timing [Forward] 111.48 ms, avg: 0.22296 ms X 500 passes Activation Operator CPU: Timing [Backward] 5.408 ms, avg: 0.010816 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator CPU: Timing [Forward] 333.439 ms, avg: 0.666878 ms X 500 passes Activation Operator CPU: Timing [Backward] 21.331 ms, avg: 0.042662 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator CPU: Timing [Forward] 3429.19 ms, avg: 6.85837 ms X 500 passes Activation Operator CPU: Timing [Backward] 286.324 ms, avg: 0.572648 ms X 500 passes GPU === Timing: 50 iterations of 10 calls, shape = [1,1,28,28] Activation Operator GPU: Timing [Forward] 1.618 ms, avg: 0.003236 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.671 ms, avg: 0.003342 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [1,3,28,28] Activation Operator GPU: Timing [Forward] 1.629 ms, avg: 0.003258 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.728 ms, avg: 0.003456 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,1,18,32] Activation Operator GPU: Timing [Forward] 1.753 ms, avg: 0.003506 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.756 ms, avg: 0.003512 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [50,3,18,32] Activation Operator GPU: Timing [Forward] 1.704 ms, avg: 0.003408 ms X 500 passes Activation Operator GPU: Timing [Backward] 1.791 ms, avg: 0.003582 ms X 500 passes Timing: 50 iterations of 10 calls, shape = [20,3,128,128] Activation Operator GPU: Timing [Forward] 2.032 ms, avg: 0.004064 ms X 500 passes Activation Operator GPU: Timing [Backward] 2.143 ms, avg: 0.004286 ms X 500 passes * lint * Trigger build * Trigger build * Negative begin and end support for csr slice (apache#8241) * negative index support for sparse slice * fix lint * getitem(int) for csr ndarray, support a[-1] * remove unneccessary argument * unittest and doc update * Preparing for 0.12.0.rc0: Final changes before RC (apache#8301) * Final changes before RC * Updates to NEWS.md * Updates * Enable smoothing in softmax operator (apache#8125) * v0.12 regression: Fix registration of children for Block (apache#8277) * Fix Block not registering children If the attribute was already set to something different than Block (e.g. None), it was not being registered. * fix if / elif for block children registration * trigger test * Add fix from apache#8152 * Add tests from apache#8152 * Revert "[CMAKE] Fix windows cmake build" (apache#8311) * Revert "Added my code signing key (apache#8293)" This reverts commit 22ab185. * Revert "[CMAKE] Fix windows cmake build (apache#8227)" This reverts commit 1c1c788. * fixed broken links. https was pointing to http for mxnet.io (apache#8300) * Update rnn.md (apache#8320) * fluent methods for missed ops (apache#8329) * update ps lite (apache#8327) * Fix unused type warning (apache#8316) * Trigger build * Trigger build * Misc fixes for sparse distributed training (apache#8345) * remove mshadow::range in init_op.h * add unit test * remove pass by ptr, add unit test for pull empty wieghts * fix range in key partition * remove wrong comment * remove change for partition * remove unused var * add int64 to arange. add checkpointing example * Fix the Readme (apache#8369) * Allow test to converge (apache#8351) * Allow test to converge * Trigger build * Trigger build * Trigger build * Update cudnn_algoreg-inl.h (apache#7988) * [Perl] emulate Python zip() for Perl (apache#8192) * [Perl] emulate Python zip() for Perl * [Perl] retool zip() uses away from the callback form * add profile option for frontend profiling to image script (apache#8171) * add profile option for frontend profiling to image script * Update image_classification.py * Update image_classification.py * Fix Typo (classification) (apache#8376) Fix a typo in the example readme. * do gtest test * add assert and do higher runs as performance test only (when performance test flag set) * Trigger build * lint * Trigger build * Sparse operator performance improvement (apache#8412) * sparse rsprsp perf improvements * Clean up * dtype default to source_array.dtype for sparse ndarrays (apache#8403) * derive default dtype/ctx from input for sparse ndarrays * add gpu tests * fix lint. add doc * remove default_ctx code * bug fix when passing dtype to array() * update doc * remove extra line * also check ctx * fix using default mean pixels (apache#8352) * fix gluon.data.RecordFileDataset (apache#8353) * upgrade MKL (apache#8378) * Lint fix (apache#8402) * Trigger build

szha requested a review from piiswrong October 5, 2017 03:45

szha force-pushed the gluon_tuple branch from 1b2185f to 4b05404 Compare October 5, 2017 03:50

piiswrong reviewed Oct 5, 2017

View reviewed changes

szha force-pushed the gluon_tuple branch from 4b05404 to 7b54584 Compare October 5, 2017 22:56

szha changed the title ~~memory save~~ gluon cached op call improvement Oct 6, 2017

szha force-pushed the gluon_tuple branch 4 times, most recently from aa2f6ff to e2f3c17 Compare October 7, 2017 21:10

szha force-pushed the gluon_tuple branch 2 times, most recently from f304210 to 085b7cb Compare October 9, 2017 17:57

szha mentioned this pull request Oct 11, 2017

add export to gluon #8212

Merged

7 tasks

szha force-pushed the gluon_tuple branch 2 times, most recently from b973920 to 4f7980e Compare October 15, 2017 05:21

szha mentioned this pull request Oct 15, 2017

v0.12 regression: Fix registration of children for Block #8277

Merged

leezu added a commit to leezu/mxnet that referenced this pull request Oct 16, 2017

Add fix from apache#8152

68e6b7b

leezu added a commit to leezu/mxnet that referenced this pull request Oct 16, 2017

Add tests from apache#8152

acedbb9

szha closed this Oct 16, 2017

szha deleted the gluon_tuple branch October 16, 2017 04:41

szha restored the gluon_tuple branch October 16, 2017 04:41

szha reopened this Oct 16, 2017

szha force-pushed the gluon_tuple branch from 4f7980e to 062893f Compare October 21, 2017 19:43

szha changed the title ~~gluon cached op call improvement~~ gluon improvement Oct 22, 2017

szha force-pushed the gluon_tuple branch from 62e6ae2 to 0182896 Compare October 22, 2017 06:32

szha force-pushed the gluon_tuple branch from 0182896 to 403c8fa Compare October 27, 2017 05:30

szha force-pushed the gluon_tuple branch 5 times, most recently from 665555d to 8f4b085 Compare November 2, 2017 02:48

fill shape after set_data

4950b8c

szha force-pushed the gluon_tuple branch from 8f4b085 to 4950b8c Compare November 2, 2017 17:16

piiswrong reviewed Nov 3, 2017

View reviewed changes

szha closed this Nov 3, 2017

szha deleted the gluon_tuple branch December 15, 2017 20:23

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

gluon improvement #8152

gluon improvement #8152

szha commented Oct 5, 2017 •

edited

Loading

piiswrong commented Oct 5, 2017

piiswrong Oct 5, 2017

szha Oct 5, 2017

szha Oct 5, 2017

piiswrong commented Oct 5, 2017

mli commented Oct 5, 2017

szha commented Oct 5, 2017 •

edited

Loading

piiswrong commented Oct 7, 2017

szha commented Oct 9, 2017

szha commented Oct 22, 2017 •

edited

Loading

cjolivier01 commented Oct 22, 2017

szha commented Oct 22, 2017 •

edited

Loading

piiswrong Nov 3, 2017

gluon improvement #8152

gluon improvement #8152

Conversation

szha commented Oct 5, 2017 • edited Loading

piiswrong commented Oct 5, 2017

piiswrong Oct 5, 2017

Choose a reason for hiding this comment

szha Oct 5, 2017

Choose a reason for hiding this comment

szha Oct 5, 2017

Choose a reason for hiding this comment

piiswrong commented Oct 5, 2017

mli commented Oct 5, 2017

szha commented Oct 5, 2017 • edited Loading

piiswrong commented Oct 7, 2017

szha commented Oct 9, 2017

szha commented Oct 22, 2017 • edited Loading

cjolivier01 commented Oct 22, 2017

szha commented Oct 22, 2017 • edited Loading

piiswrong Nov 3, 2017

Choose a reason for hiding this comment

szha commented Oct 5, 2017 •

edited

Loading

szha commented Oct 5, 2017 •

edited

Loading

szha commented Oct 22, 2017 •

edited

Loading

szha commented Oct 22, 2017 •

edited

Loading