Benchmarking - handle all NaN signal for WFDB write_waveform format

[briangow]

If a requested section of a signal was composed of all NaNs, the WFDB write_waveform function would fail with this error:

Traceback (most recent call last):
  File "/chorus_waveform/./waveform_benchmark.py", line 6, in <module>
    waveform_benchmark.__main__.main()
  File "/chorus_waveform/waveform_benchmark/__main__.py", line 110, in main
    format_list, waveform_list, test_list, result_list = run_benchmarks(input_record=record,
                                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/chorus_waveform/waveform_benchmark/benchmark.py", line 114, in run_benchmarks
    fmt().write_waveforms(path, waveforms)
  File "/chorus_waveform/waveform_benchmark/formats/wfdb.py", line 48, in write_waveforms
    sig_baseline = round(-sig_gain * (sample_min + sample_max) / 2)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: cannot convert float NaN to integer

This was occurring with this waveform: mimic3wdb/1.0/36/3654093/3654093. This PR first checks to see if the requested signal is composed of all NaNs. If it is, sig_baseline is set to 0. If it isn't, the baseline is set as was being done previously.

[bemoody]

Thanks! Yeah, that should work, but this ought to be more efficient:

            # Determine the minimum and maximum non-NaN sample values.
            # (nanmin and nanmax will give a warning if all values are NaN.)
            sample_min = numpy.fmin.reduce(sig_samples)
            sample_max = numpy.fmax.reduce(sig_samples)
            if numpy.isnan(sample_min):
                sig_baseline = 0
            else:
                sig_baseline = round(-sig_gain * (sample_min + sample_max) / 2)

Alternatively, I think numpy.isnan(sig_samples).all() would be better than all(numpy.isnan(sig_samples)).

[briangow]

Thanks @bemoody , I've updated with your suggestion. I'll merge this now.