[MXNET-349] Histogram Operator #10931

haojin2 · 2018-05-14T08:18:58Z

Description

This PR implements histogram operator.

Checklist

Essentials

Changes are complete (i.e. I finished coding on this PR)
All changes have test coverage:
Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
Code is well-documented:
For user-facing API changes, API doc string has been updated.
For new C++ functions in header files, their functionalities and arguments are documented.
For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

Implementation of internal _histogram operator
User-facing Python interface
Unit tests

Comments

The user-facing interface resembles basic behaviors of Numpy's histogram: https://docs.scipy.org/doc/numpy-1.14.0/reference/generated/numpy.histogram.html.
Three versions are supported:
histogram(x, bins=7) - divided into 7 equal bins in the range of [x.min, x.max]
histogram(x, bins=7, range=(a,b)) - divided into 7 equal bins in the range of [a, b], anything outside the range is ignored
histogram(x, bins=NDArray) - bin bounds are specified by the bins NDArray.

piiswrong · 2018-05-14T17:26:51Z

python/mxnet/ndarray/ndarray.py

+        return _internal._histogram(data=a, bins=bins)
+    elif isinstance(bins, int):
+        if range_ is None:
+            range_ = (float(a.min().asnumpy()[0]), float(a.max().asnumpy()[0]))


don't do this in the frontend.

piiswrong · 2018-05-14T17:27:22Z

python/mxnet/ndarray/ndarray.py

+    elif isinstance(bins, int):
+        if range_ is None:
+            range_ = (float(a.min().asnumpy()[0]), float(a.max().asnumpy()[0]))
+        return _internal._histogram(data=a, bins=array([]), bin_cnt=bins, range=range_)


why make an invalid array/

bins is not used in this version of _histogram, but it's one of the inputs so I need a mock array.

You can define a mode variable in HistogramParam for distinguishing the cases of 1 and 2 ndarray inputs. When register the op, use set_num_inputs([](const NodeAttrs& attrs) to define the number of inputs based upon the mode.
See example:
https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/convolution.cc#L471

piiswrong · 2018-05-14T17:36:36Z

src/operator/tensor/histogram.cc

+.add_argument("bins", "NDArray-or-Symbol", "Input ndarray")
+.add_arguments(HistogramParam::__FIELDS__());
+
+NNVM_REGISTER_OP(_backward_histogram)


Historgram op is not differentiable. Don't register backward op.

Sure, will get rid of it.

reminisce · 2018-05-14T22:21:47Z

python/mxnet/ndarray/ndarray.py

@@ -3740,3 +3740,32 @@ def empty(shape, ctx=None, dtype=None):
    if dtype is None:
        dtype = mx_real_t
    return NDArray(handle=_new_alloc_handle(shape, ctx, False, dtype))
+
+
+def histogram(a, bins=10, range_=None):


Why underscore suffix for range?

pylint complains if we override the keyword range

@piiswrong numpy.histogram has parameter range. Should we disable pylint check here to make mx.nd.histogram parameter naming consistent with numpy?

reminisce · 2018-05-14T22:22:29Z

python/mxnet/ndarray/ndarray.py

+    """
+
+    # pylint: disable= no-member, protected-access
+    if isinstance(bins, NDArray):


Need to check whether bins is a 1d array?

That's checked inside the op

reminisce · 2018-05-14T22:23:09Z

python/mxnet/ndarray/ndarray.py

+    # pylint: disable= no-member, protected-access
+    if isinstance(bins, NDArray):
+        return _internal._histogram(data=a, bins=bins)
+    elif isinstance(bins, int):


Replace int with integer_types.

reminisce · 2018-05-14T22:26:29Z

tests/python/unittest/test_operator.py

+    def f(x, bins=10, range=None):
+        return np.histogram(x, bins, range=range)
+
+    shape = (3, 3, 3)


Randomize shape and test ndarrays from 1D-5D.

reminisce · 2018-05-14T22:29:01Z

tests/python/unittest/test_operator.py

+    x = rand_ndarray(shape, stype='default')
+    mx_bins = mx.nd.array([-1.0, 0.5, 2.0, 4.5, 50.0])
+    np_bins = mx_bins.asnumpy()
+    bin_cnt = 5


Should test more settings of bin_cnt.

reminisce · 2018-05-14T22:34:13Z

python/mxnet/ndarray/ndarray.py

+    elif isinstance(bins, int):
+        if range_ is None:
+            range_ = (float(a.min().asnumpy()[0]), float(a.max().asnumpy()[0]))
+        return _internal._histogram(data=a, bins=array([]), bin_cnt=bins, range=range_)


You can define a mode variable in HistogramParam for distinguishing the cases of 1 and 2 ndarray inputs. When register the op, use set_num_inputs([](const NodeAttrs& attrs) to define the number of inputs based upon the mode.
See example:
https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/convolution.cc#L471

reminisce · 2018-05-14T22:38:32Z

src/operator/tensor/histogram-inl.h

+    SHAPE_ASSIGN_CHECK(*out_attrs, 1, in_attrs->at(1));
+  }
+
+  return out_attrs->at(0).ndim() == 1U && out_attrs->at(0).Size() != 0U &&


Call shape_is_none to simplify the code. ndim == 1 should have been guaranteed before this line.

reminisce · 2018-05-14T22:39:12Z

src/operator/tensor/histogram-inl.h

+
+  TYPE_ASSIGN_CHECK(*out_attrs, 0, mshadow::kInt64);
+  TYPE_ASSIGN_CHECK(*out_attrs, 1, in_attrs->at(0));
+  return out_attrs->at(0) != -1 && out_attrs->at(1) != -1;


Call !type_is_none(out_attrs->at(i))

haojin2 · 2018-06-02T04:47:00Z

@reminisce @piiswrong @anirudh2290 @rahul003 Unit tests added, this is ready, please give a review if you have time. Thanks!

piiswrong · 2018-06-02T21:43:14Z

python/mxnet/ndarray/ndarray.py

+            res, bin_bounds = np.histogram(a.asnumpy(), bins=bins)
+            return array(res), array(bin_bounds)
+        return _internal._histogram(data=a, bin_cnt=bins, range=range)
+    return None


why return None? Shouldn't this cause an error?

Do you mean that I should throw an exception?

Added an exception for the illegal input case.

reminisce · 2018-06-05T22:59:04Z

src/operator/tensor/histogram.cc

+.set_attr<nnvm::FInferShape>("FInferShape", HistogramOpShape)
+.set_attr<nnvm::FInferType>("FInferType", HistogramOpType)
+.set_attr<FCompute>("FCompute<cpu>", HistogramOpForward<cpu>)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",


No need to register this if it's empty.

reminisce · 2018-06-05T22:59:13Z

src/operator/tensor/histogram.cc

+  }
+}
+
+template<typename cpu>


Should this be a template specialization? Same for other places.

template<> void HistogramForwardImpl<cpu>()

reminisce · 2018-06-06T05:10:58Z

src/operator/tensor/histogram.cc

+}
+
+template<>
+void HistogramForwardImpl<cpu>(const OpContext& ctx,


Why change this back to using OpContext? If using Stream for dispatching, you can get rid of the templates.

I'm not using OpContext, I'm just getting rid of the stream argument here.

haojin2 · 2018-06-07T00:15:43Z

@piiswrong Should be ready for merge, please take a look when you have time, thanks!

piiswrong · 2018-06-12T00:25:36Z

So symbolic is not supported?

haojin2 · 2018-06-12T00:28:21Z

@piiswrong There's no backward pass for this function, I wonder if symbolic is necessary in this case?

haojin2 · 2018-06-22T17:55:25Z

@piiswrong symbol is added for histogram.

* implementation of histogram operator * address code reviews and code re-design * add exception for invalid inputs * address code reviews * add symbol and symbolic forward check for histogram

haojin2 requested a review from szha as a code owner May 14, 2018 08:18

haojin2 force-pushed the histogram branch 2 times, most recently from a1bf838 to 49a9ae4 Compare May 14, 2018 08:31

piiswrong reviewed May 14, 2018

View reviewed changes

haojin2 force-pushed the histogram branch 3 times, most recently from 7c44e7f to b17585a Compare May 14, 2018 20:50

reminisce suggested changes May 14, 2018

View reviewed changes

haojin2 force-pushed the histogram branch from b17585a to 652c702 Compare May 15, 2018 00:21

haojin2 force-pushed the histogram branch 4 times, most recently from 26a9d38 to c2a9f1d Compare May 30, 2018 20:38

haojin2 changed the title ~~[MXNET-349] [WIP] Histogram Operator~~ [MXNET-349] Histogram Operator May 30, 2018

haojin2 force-pushed the histogram branch 4 times, most recently from 240eecc to 55d6ccb Compare June 2, 2018 01:38

piiswrong suggested changes Jun 2, 2018

View reviewed changes

haojin2 force-pushed the histogram branch 2 times, most recently from 39fe74e to c53e477 Compare June 4, 2018 23:47

reminisce reviewed Jun 5, 2018

View reviewed changes

haojin2 force-pushed the histogram branch from c53e477 to 7ecc23d Compare June 6, 2018 00:09

reminisce reviewed Jun 6, 2018

View reviewed changes

haojin2 force-pushed the histogram branch 3 times, most recently from 8689ebd to ee38e0c Compare June 6, 2018 20:39

reminisce approved these changes Jun 6, 2018

View reviewed changes

haojin2 force-pushed the histogram branch from ee38e0c to c9ad77c Compare June 13, 2018 04:54

haojin2 force-pushed the histogram branch from c9ad77c to c2619fc Compare June 22, 2018 17:54

haojin2 force-pushed the histogram branch 2 times, most recently from 8bc8e18 to 4595fd9 Compare June 23, 2018 16:24

Hao Jin added 5 commits June 25, 2018 18:03

implementation of histogram operator

766da6a

address code reviews and code re-design

9bcc18a

add exception for invalid inputs

3035cc3

address code reviews

0898151

add symbol and symbolic forward check for histogram

9cff47b

haojin2 force-pushed the histogram branch from 4595fd9 to 9cff47b Compare June 25, 2018 18:03

piiswrong merged commit ed7e360 into apache:master Jun 25, 2018

haojin2 deleted the histogram branch March 25, 2019 08:14

[MXNET-349] Histogram Operator #10931

[MXNET-349] Histogram Operator #10931

Conversation

haojin2 commented May 14, 2018

Description

Checklist

Essentials

Changes

Comments

Choose a reason for hiding this comment

piiswrong May 14, 2018 • edited Loading

Choose a reason for hiding this comment

haojin2 May 14, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

haojin2 commented Jun 2, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

haojin2 commented Jun 7, 2018

piiswrong commented Jun 12, 2018

haojin2 commented Jun 12, 2018

haojin2 commented Jun 22, 2018

piiswrong May 14, 2018 •

edited

Loading

haojin2 May 14, 2018 •

edited

Loading