Bug in np.histogram when min == max and values are >= 2**53

The following call to histogram fails:

In [5]: np.histogram([2**53])
/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py:566: RuntimeWarning: divide by zero encountered in double_scalars
  norm = bins / (mx - mn)
/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py:591: RuntimeWarning: invalid value encountered in multiply
  tmp_a *= norm
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-5-41f77d7d86af> in <module>()
----> 1 np.histogram([2**53])

/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py in histogram(a, bins, range, normed, weights, density)
    598             # The index computation is not guaranteed to give exactly
    599             # consistent results within ~1 ULP of the bin edges.
--> 600             decrement = tmp_a_data < bin_edges[indices]
    601             indices[decrement] -= 1
    602             # The last bin includes the right edge. The other bins do not.

IndexError: index -9223372036854775808 is out of bounds for axis 1 with size 11

whereas this succeeds:

In [6]: np.histogram([2**53-1])
Out[6]: 
(array([0, 0, 0, 0, 0, 0, 1, 0, 0, 0]),
 array([  9.00719925e+15,   9.00719925e+15,   9.00719925e+15,
          9.00719925e+15,   9.00719925e+15,   9.00719925e+15,
          9.00719925e+15,   9.00719925e+15,   9.00719925e+15,
          9.00719925e+15,   9.00719925e+15]))

The threshold for triggering the bug appears to be 2**53, but this only happens when the min of the array is the same as the max:

In [7]: np.histogram([2**53, 2**53])
/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py:566: RuntimeWarning: divide by zero encountered in double_scalars
  norm = bins / (mx - mn)
/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py:591: RuntimeWarning: invalid value encountered in multiply
  tmp_a *= norm
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-7-f74f7a6cf401> in <module>()
----> 1 np.histogram([2**53, 2**53])

/Users/tom/miniconda3/envs/dev/lib/python3.6/site-packages/numpy/lib/function_base.py in histogram(a, bins, range, normed, weights, density)
    598             # The index computation is not guaranteed to give exactly
    599             # consistent results within ~1 ULP of the bin edges.
--> 600             decrement = tmp_a_data < bin_edges[indices]
    601             indices[decrement] -= 1
    602             # The last bin includes the right edge. The other bins do not.

IndexError: index -9223372036854775808 is out of bounds for axis 1 with size 11

When the min and the max are different, things are fine:

In [8]: np.histogram([2**53, 2**54])
Out[8]: 
(array([1, 0, 0, 0, 0, 0, 0, 0, 0, 1]),
 array([  9.00719925e+15,   9.90791918e+15,   1.08086391e+16,
          1.17093590e+16,   1.26100790e+16,   1.35107989e+16,
          1.44115188e+16,   1.53122387e+16,   1.62129587e+16,
          1.71136786e+16,   1.80143985e+16]))

[sanjok-bless]

I have the same issue with a big double number, reproducible code:

double_numbers = np.array([1e20] * 20)
np.histogram(double_numbers, bins=100)

I get an error:

IndexError                                Traceback (most recent call last)
<ipython-input-2-51958daa5f2c> in <module>()
      1 double_numbers = np.array([1e20] * 20)
----> 2 np.histogram(double_numbers, bins=100)

/home/oleksandr/.pyenv/versions/2.7.8/envs/dr2.7.8/lib/python2.7/site-packages/numpy/lib/histograms.pyc in histogram(a, bins, range, normed, weights, density)
    764             # The index computation is not guaranteed to give exactly
    765             # consistent results within ~1 ULP of the bin edges.
--> 766             decrement = tmp_a < bin_edges[indices]
    767             indices[decrement] -= 1
    768             # The last bin includes the right edge. The other bins do not.

IndexError: index -9223372036854775808 is out of bounds for axis 1 with size 101

I use numpy 1.15.1 and see the bug exists at least from 1.10 version. Are there any plans to fix this?

[collinmccarthy]

I’m getting this same error with the following: np.histogram(np.nan_to_num(np.inf))

This makes it difficult to create a histogram when some values are +/- infinity, which happens a lot when logging gradient information in deep learning frameworks like PyTorch and Tensorflow.

Are there still plans to fix this issue? If not what should I set these large values to (sometimes they’re in 32 or 64-bit arrays) to avoid this? Thank you.