Description

Hi, there.
I fix the wrong result of sum(inf, inf) and mean(inf, inf).

Test Case:

import mxnet as mx
import numpy as np

def test(x):
    print('data', x.asnumpy())
    print('mean/sum', mx.nd.mean(x).asnumpy(), mx.nd.sum(x).asnumpy())
    print('argmin/argmax', mx.nd.argmin(x).asnumpy(), mx.nd.argmax(x).asnumpy())
    print('min/max', mx.nd.min(x).asnumpy(), mx.nd.max(x).asnumpy())
    print('-----')

x = mx.nd.array([np.inf, np.inf, 1])
test(x)

x = mx.nd.array([-np.inf, -np.inf, 1])
test(x)

x = mx.nd.array([np.inf, -np.inf, 1])
test(x)

x = mx.nd.array([np.nan, -np.inf, 1])
test(x)

x = mx.nd.array([np.nan, np.nan])
test(x)

x = mx.nd.array([np.nan, 1])
test(x)

The wrong result in mxnet_mkl-1.6.0b20191015-py2.py3-none-manylinux1_x86_64

data [inf inf  1.]
mean/sum [nan] [nan]
argmin/argmax [2.] [0.]
min/max [1.] [inf]
-----
data [-inf -inf   1.]
mean/sum [nan] [nan]
argmin/argmax [0.] [2.]
min/max [-inf] [1.]
-----
data [ inf -inf   1.]
mean/sum [nan] [nan]
argmin/argmax [1.] [0.]
min/max [-inf] [inf]
-----
data [ nan -inf   1.]
mean/sum [nan] [nan]
argmin/argmax [1.] [2.]
min/max [-inf] [1.]
-----
data [nan nan]
mean/sum [nan] [nan]
argmin/argmax [0.] [0.]
min/max [inf] [-inf]
-----
data [nan  1.]
mean/sum [nan] [nan]
argmin/argmax [1.] [1.]
min/max [1.] [1.]

The correct result in this PR:

data [inf inf  1.]
mean/sum [inf] [inf]
argmin/argmax [2.] [0.]
min/max [1.] [inf]
-----
data [-inf -inf   1.]
mean/sum [-inf] [-inf]
argmin/argmax [0.] [2.]
min/max [-inf] [1.]
-----
data [ inf -inf   1.]
mean/sum [nan] [nan]
argmin/argmax [1.] [0.]
min/max [-inf] [inf]
-----
data [ nan -inf   1.]
mean/sum [nan] [nan]
argmin/argmax [0.] [0.]
min/max [nan] [nan]
-----
data [nan nan]
mean/sum [nan] [nan]
argmin/argmax [0.] [0.]
min/max [nan] [nan]
-----
data [nan  1.]
mean/sum [nan] [nan]
argmin/argmax [0.] [0.]
min/max [nan] [nan]

If we modify the test function,

def test(x):
    x = x.asnumpy()
    print('data', x)
    print('mean/sum', np.mean(x), np.sum(x))
    print('argmin/argmax', np.argmin(x), np.argmax(x))
    print('min/max', np.min(x), np.max(x))
    print('-----')

Here is the result of NumPy.

data [inf inf  1.]
mean/sum inf inf
argmin/argmax 2 0
min/max 1.0 inf
-----
data [-inf -inf   1.]
mean/sum -inf -inf
argmin/argmax 0 2
min/max -inf 1.0
-----
data [ inf -inf   1.]
mean/sum nan nan
argmin/argmax 1 0
min/max -inf inf
-----
data [ nan -inf   1.]
mean/sum nan nan
argmin/argmax 0 0
min/max nan nan
-----
data [nan nan]
mean/sum nan nan
argmin/argmax 0 0
min/max nan nan
-----
data [nan  1.]
mean/sum nan nan
argmin/argmax 0 0
min/max nan nan
-----

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$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

• Add isnan_typed and isinf_typed in mshadow
• Remove isnan_typed in src/operator/mshadow_op.h
• Update mshadow/extension/reduce_with_axis.h to support NaN
• Add relative python testcases
• Update Julia testcase, which is consistent with Julia built-in functions argmin and argmax.