mne.find_event()'s uint_cast needs to be more flexible

Describe the bug

I am currently working with Cam-CAN data, which were acquired using a Neuromag MEG system. According to the mne.find_events() docs, there may be issues when reading events created on such a system if STI016 was active. In the data I have, there is no STI016; however, my stimulus channel of interest, STI101, yields spurious events, even when passing the recommended uint_cast=True parameter to mne.find_events. I found that this param causes a type cast to uint16; yet, it seems my data requires to be cast to uint8 instead.

Steps to reproduce

I have prepared a raw FIF file that only contains the STI101 channel: https://drive.google.com/open?id=1rpgvxNSyWti6_VoAaFaCqsKgYAQ2wWUh

import numpy as np
import mne

infile = '/tmp/STI101_raw.fif'
sti101_raw = mne.io.read_raw_fif(infile, verbose=False)

events = mne.find_events(sti101_raw, verbose=False)
print(np.unique(events[:, 2]))

events_uint_cast = mne.find_events(sti101_raw, verbose=False,
                                   uint_cast=True)
print(np.unique(events_uint_cast[:, 2]))

# Handle the data manually.
data = sti101_raw.get_data().squeeze()
print(np.unique(data))
print(np.unique(data.astype(np.uint8)))

Expected results

[    1     2     3     4     5     6     7 ]
[    1     2     3     4     5     6     7 ]
[-3.2768e+04 -3.2764e+04  0.0000e+00  1.0000e+00  2.0000e+00  3.0000e+00
  4.0000e+00  5.0000e+00  6.0000e+00  7.0000e+00]
[0 1 2 3 4 5 6 7]

Actual results

[    1     2     3     4     5     6     7 32768]
[    1     2     3     4     5     6     7 32768 32772]
[-3.2768e+04 -3.2764e+04  0.0000e+00  1.0000e+00  2.0000e+00  3.0000e+00
  4.0000e+00  5.0000e+00  6.0000e+00  7.0000e+00]
[0 1 2 3 4 5 6 7]

As you can see, only when casting to uint8, I get the correct results (last line)

[agramfort]

go for it. Indeed it’s a frequent pb.

[u01ai11]

Not that this impacts the solution, but you might be interested in where these spurious events are coming from. It may be an issue with set-ups which have devices (e.g. button boxes) share channels with software triggers – which impacts the calculation of STI101. I think this is done to preserve parallel port compatibility for different presentation software. For instance the Cambridge Neuromag setup has this quirk detailed here: https://imaging.mrc-cbu.cam.ac.uk/meg/Triggers .