some big numbers can not be assigned correctly with ndarray, it is a big bug???

[starimpact]

code:

  1 import sys
  2 sys.path.insert(0, '/home/mingzhang/work/dmlc/mxnet_v0.8.0.python')
  3
  4 import mxnet as mx
  5 import numpy as np
  6
  7 a = mx.nd.zeros((4, 1), dtype=np.int32)
  8 a[1,0] = int(16800001)
  9 a[2,0] = int(16800002)
 10 b = a.asnumpy()
 11 print b.T, b.dtype
 12 c = a-1
 13 b = c.asnumpy()
 14 print b.T, b.dtype

output:

[[       0 16800000 16800002        0]] int32
[[      -1 16799999 16800001       -1]] int32

16800001 can not be assigned correctly!
the problem is also in mxnet v1.0 and v1.2

[starimpact]

@tqchen @piiswrong

[szha]

reproducible on master

[reminisce]

This is a problem of using float32 by passing big integers to the backend from frontend. In this case, 16800001 becomes 1.68e+07. That's why 1 is missing. I think we can add a int field in the _slice_assign_scalar operator and use that to pass integers. @piiswrong What do you think?

[apeforest]

@reminisce I have fixed this issue by using a double data type to hold large integers. Please help to review the fix. Thanks.

[apeforest]

@sandeep-krishnamurthy @nswamy This issue is fixed. Please close it. Thanks!