Overflow when running reductions on float16 columns in pandas Series

[jsleroux]

When running reductions on dataframe columns of dtype float16, we ran into a surprising behaviour:

import pandas as pd
df = pd.DataFrame(data=[1000]*100, columns=['A'], dtype='float16')

mean = df['A'].mean()
sum = df['A'].sum()
print('Mean: {}, sum: {}'.format(mean, sum))

# Both of these should be true, but they're not, returning inf in pandas 0.23.4
assert mean == 1000
assert sum == 1000 * 100

After investigation, we found that the accumulator used in (for example) mean() and sum() is not big enough and it eventually overflows. In both nansum() and nanmean() functions (to which sum() and mean() delegate their work), when the data type is 'float', the accumulator is downcasted to the original dtype of the data:

• nansum:

  pandas/pandas/core/nanops.py

  Line 333 in af7b0ba

  dtype_sum = dtype

• nanmean:

  pandas/pandas/core/nanops.py

  Line 352 in af7b0ba

  dtype_sum = dtype

In our case, because the original dtype is float16, the accumulator is downcasted to float16, for example in nansum():

    dtype_sum = dtype_max
    if is_float_dtype(dtype):
        dtype_sum = dtype

(Similar code in nanmean.)

However, pandas did not always behave like that. The current behaviour was added in the following commits, to solve other bugs:

• sum(): 73f25b1
• mean(): 3896e5e

Also, for the "int" codepaths, the accumulator is never downcast and always set to float64. It is only for the "float" cases that the size of the accumulator is set to be the same as the dtype of the column.

We'd be willing to submit a pull request, but we're not sure what the best fix here would be. Should we just always have a float64 accumulator in these functions for the float cases, instead of downcasting it? If not, what would a good fix look like?

By the way, things are a bit different in numpy, with the accumulator being set to float64 in more cases, and with the option for users to specify the dtype of the accumulator (and at the same time the output). Having the same option in pandas would have allowed us to at least work around this, by requesting a float64 accumulator. What there a decision made in pandas not to offer a dtype argument to sum, etc. like numpy does? Otherwise, we could implement that also in the pull request.

(Cc @chrish42)

[jreback]

float16 has no support whatsoever

that said would take a patch that upcasts accumulators to a bigger dtype / uses float64 for float type ops if < float64

[Zac-HD]

Tag as good first issue?

[chrish42]

@Zac-HD Agree, but it would already be @jsleroux's first issue, and I've been coaching him a bit for this. We're discussed creating the pull request at the office last Friday. Can we hold off that tag for a little bit, while he works on this? :-)

[Zac-HD]

Oh, that's fantastic! I was just thinking of how to help someone find this, but for @jsleroux to fix it too would be great 🎉