""" Camera extractor functions
This module handles extraction of camera timestamps for both Bpod and FPGA.
"""
import logging
from functools import partial
import cv2
import numpy as np
import matplotlib.pyplot as plt
from iblutil.util import range_str
import neurodsp.utils as dsp
from ibllib.plots import squares, vertical_lines
from ibllib.io.video import assert_valid_label, VideoStreamer
from iblutil.numerical import within_ranges
from ibllib.io.extractors.base import get_session_extractor_type
from ibllib.io.extractors.ephys_fpga import get_sync_fronts, get_sync_and_chn_map
import ibllib.io.raw_data_loaders as raw
from ibllib.io.extractors.base import (
BaseBpodTrialsExtractor,
BaseExtractor,
run_extractor_classes,
_get_task_types_json_config
)
_logger = logging.getLogger(__name__)
[docs]def get_video_length(video_path):
"""
Returns video length
:param video_path: A path to the video
:return:
"""
is_url = isinstance(video_path, str) and video_path.startswith('http')
cap = VideoStreamer(video_path).cap if is_url else cv2.VideoCapture(str(video_path))
assert cap.isOpened(), f'Failed to open video file {video_path}'
length = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
cap.release()
return length
[docs]class CameraTimestampsFPGA(BaseExtractor):
def __init__(self, label, session_path=None):
super().__init__(session_path)
self.label = assert_valid_label(label)
self.save_names = f'_ibl_{label}Camera.times.npy'
self.var_names = f'{label}_camera_timestamps'
self._log_level = _logger.level
_logger.setLevel(logging.DEBUG)
def __del__(self):
_logger.setLevel(self._log_level)
def _extract(self, sync=None, chmap=None, video_path=None,
display=False, extrapolate_missing=True, **kwargs):
"""
The raw timestamps are taken from the FPGA. These are the times of the camera's frame TTLs.
If the pin state file exists, these timestamps are aligned to the video frames using the
audio TTLs. Frames missing from the embedded frame count are removed from the timestamps
array.
If the pin state file does not exist, the left and right camera timestamps may be aligned
using the wheel data.
:param sync: dictionary 'times', 'polarities' of fronts detected on sync trace.
:param chmap: dictionary containing channel indices. Default to constant.
:param video_path: an optional path for fetching the number of frames. If None,
the video is loaded from the session path. If an int is provided this is taken to be
the total number of frames.
:param display: if True, the audio and GPIO fronts are plotted.
:param extrapolate_missing: if True, any missing timestamps at the beginning and end of
the session are extrapolated based on the median frame rate, otherwise they will be NaNs.
:return: a numpy array of camera timestamps
"""
fpga_times = extract_camera_sync(sync=sync, chmap=chmap)
count, (*_, gpio) = raw.load_embedded_frame_data(self.session_path, self.label)
raw_ts = fpga_times[self.label]
if video_path is None:
filename = f'_iblrig_{self.label}Camera.raw.mp4'
video_path = self.session_path.joinpath('raw_video_data', filename)
# Permit the video path to be the length for development and debugging purposes
length = (video_path if isinstance(video_path, int) else get_video_length(video_path))
_logger.debug(f'Number of video frames = {length}')
if gpio is not None and gpio['indices'].size > 1:
_logger.info('Aligning to audio TTLs')
# Extract audio TTLs
audio = get_sync_fronts(sync, chmap['audio'])
_, ts = raw.load_camera_ssv_times(self.session_path, self.label)
try:
"""
NB: Some of the audio TTLs occur very close together, and are therefore not
reflected in the pin state. This function removes those. Also converts frame
times to FPGA time.
"""
gpio, audio, ts = groom_pin_state(gpio, audio, ts, display=display)
"""
The length of the count and pin state are regularly longer than the length of
the video file. Here we assert that the video is either shorter or the same
length as the arrays, and we make an assumption that the missing frames are
right at the end of the video. We therefore simply shorten the arrays to match
the length of the video.
"""
if count.size > length:
count = count[:length]
else:
assert length == count.size, 'fewer counts than frames'
raw_ts = fpga_times[self.label]
assert raw_ts.shape[0] > 0, 'no timestamps found in channel indicated for ' \
f'{self.label} camera'
return align_with_audio(raw_ts, audio, gpio, count,
display=display,
extrapolate_missing=extrapolate_missing)
except AssertionError as ex:
_logger.critical('Failed to extract using audio: %s', ex)
# If you reach here extracting using audio TTLs was not possible
_logger.warning('Alignment by wheel data not yet implemented')
if length < raw_ts.size:
df = raw_ts.size - length
_logger.info(f'Discarding first {df} pulses')
raw_ts = raw_ts[df:]
return raw_ts
[docs]class CameraTimestampsCamlog(BaseExtractor):
def __init__(self, label, session_path=None):
super().__init__(session_path)
self.label = assert_valid_label(label)
self.save_names = f'_ibl_{label}Camera.times.npy'
self.var_names = f'{label}_camera_timestamps'
self._log_level = _logger.level
_logger.setLevel(logging.DEBUG)
def __del__(self):
_logger.setLevel(self._log_level)
def _extract(self, sync=None, chmap=None, video_path=None,
display=False, extrapolate_missing=True, **kwargs):
fpga_times = extract_camera_sync(sync=sync, chmap=chmap)
video_frames = get_video_length(self.session_path.joinpath('raw_video_data', f'_iblrig_{self.label}Camera.raw.mp4'))
raw_ts = fpga_times[self.label]
# For left camera sometimes we have one extra pulse than video frame
if (raw_ts.size - video_frames) == 1:
_logger.warning(f'One extra sync pulse detected for {self.label} camera')
raw_ts = raw_ts[:-1]
elif (raw_ts.size - video_frames) == -1:
_logger.warning(f'One extra video frame detected for {self.label} camera')
med_time = np.median(np.diff(raw_ts))
raw_ts = np.r_[raw_ts, np.array([raw_ts[-1] + med_time])]
assert video_frames == raw_ts.size, f'dimension mismatch between video frames and TTL pulses for {self.label} camera' \
f'by {np.abs(video_frames - raw_ts.size)} frames'
return raw_ts
[docs]class CameraTimestampsBpod(BaseBpodTrialsExtractor):
"""
Get the camera timestamps from the Bpod
The camera events are logged only during the events not in between, so the times need
to be interpolated
"""
save_names = '_ibl_leftCamera.times.npy'
var_names = 'left_camera_timestamps'
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self._log_level = _logger.level
_logger.setLevel(logging.DEBUG)
def __del__(self):
_logger.setLevel(self._log_level)
def _extract(self, video_path=None, display=False, extrapolate_missing=True, **kwargs):
"""
The raw timestamps are taken from the Bpod. These are the times of the camera's frame TTLs.
If the pin state file exists, these timestamps are aligned to the video frames using the
audio TTLs. Frames missing from the embedded frame count are removed from the timestamps
array.
If the pin state file does not exist, the left camera timestamps may be aligned using the
wheel data.
:param video_path: an optional path for fetching the number of frames. If None,
the video is loaded from the session path. If an int is provided this is taken to be
the total number of frames.
:param display: if True, the audio and GPIO fronts are plotted.
:param extrapolate_missing: if True, any missing timestamps at the beginning and end of
the session are extrapolated based on the median frame rate, otherwise they will be NaNs.
:return: a numpy array of camera timestamps
"""
raw_ts = self._times_from_bpod()
count, (*_, gpio) = raw.load_embedded_frame_data(self.session_path, 'left')
if video_path is None:
filename = '_iblrig_leftCamera.raw.mp4'
video_path = self.session_path.joinpath('raw_video_data', filename)
# Permit the video path to be the length for development and debugging purposes
length = video_path if isinstance(video_path, int) else get_video_length(video_path)
_logger.debug(f'Number of video frames = {length}')
# Check if the GPIO is usable for extraction. GPIO is None if the file does not exist,
# is empty, or contains only one value (i.e. doesn't change)
if gpio is not None and gpio['indices'].size > 1:
_logger.info('Aligning to audio TTLs')
# Extract audio TTLs
_, audio = raw.load_bpod_fronts(self.session_path, data=self.bpod_trials)
_, ts = raw.load_camera_ssv_times(self.session_path, 'left')
"""
There are many audio TTLs that are for some reason missed by the GPIO. Conversely
the last GPIO doesn't often correspond to any audio TTL. These will be removed.
The drift appears to be less severe than the FPGA, so when assigning TTLs we'll take
the nearest TTL within 500ms. The go cue TTLs comprise two short pulses ~3ms apart.
We will fuse any TTLs less than 5ms apart to make assignment more accurate.
"""
try:
gpio, audio, ts = groom_pin_state(gpio, audio, ts, take='nearest',
tolerance=.5, min_diff=5e-3, display=display)
if count.size > length:
count = count[:length]
else:
assert length == count.size, 'fewer counts than frames'
return align_with_audio(raw_ts, audio, gpio, count,
extrapolate_missing, display=display)
except AssertionError as ex:
_logger.critical('Failed to extract using audio: %s', ex)
# If you reach here extracting using audio TTLs was not possible
_logger.warning('Alignment by wheel data not yet implemented')
# Extrapolate at median frame rate
n_missing = length - raw_ts.size
if n_missing > 0:
_logger.warning(f'{n_missing} fewer Bpod timestamps than frames; '
f'{"extrapolating" if extrapolate_missing else "appending nans"}')
frate = np.median(np.diff(raw_ts))
to_app = ((np.arange(n_missing, ) + 1) / frate + raw_ts[-1]
if extrapolate_missing
else np.full(n_missing, np.nan))
raw_ts = np.r_[raw_ts, to_app] # Append the missing times
elif n_missing < 0:
_logger.warning(f'{abs(n_missing)} fewer frames than Bpod timestamps')
_logger.info(f'Discarding first {abs(n_missing)} pulses')
raw_ts = raw_ts[abs(n_missing):]
return raw_ts
def _times_from_bpod(self):
ntrials = len(self.bpod_trials)
cam_times = []
n_frames = 0
n_out_of_sync = 0
missed_trials = []
for ind in np.arange(ntrials):
# get upgoing and downgoing fronts
pin = np.array(self.bpod_trials[ind]['behavior_data']
['Events timestamps'].get('Port1In'))
pout = np.array(self.bpod_trials[ind]['behavior_data']
['Events timestamps'].get('Port1Out'))
# some trials at startup may not have the camera working, discard
if np.all(pin) is None:
missed_trials.append(ind)
continue
# if the trial starts in the middle of a square, discard the first downgoing front
if pout[0] < pin[0]:
pout = pout[1:]
# same if the last sample is during an upgoing front,
# always put size as it happens last
pin = pin[:pout.size]
frate = np.median(np.diff(pin))
if ind > 0:
"""
assert that the pulses have the same length and that we don't miss frames during
the trial, the refresh rate of bpod is 100us
"""
test1 = np.all(np.abs(1 - (pin - pout) / np.median(pin - pout)) < 0.1)
test2 = np.all(np.abs(np.diff(pin) - frate) <= 0.00011)
if not all([test1, test2]):
n_out_of_sync += 1
# grow a list of cam times for ech trial
cam_times.append(pin)
n_frames += pin.size
if missed_trials:
_logger.debug('trial(s) %s missing TTL events', range_str(missed_trials))
if n_out_of_sync > 0:
_logger.warning(f"{n_out_of_sync} trials with bpod camera frame times not within"
f" 10% of the expected sampling rate")
t_first_frame = np.array([c[0] for c in cam_times])
t_last_frame = np.array([c[-1] for c in cam_times])
frate = 1 / np.nanmedian(np.array([np.median(np.diff(c)) for c in cam_times]))
intertrial_duration = t_first_frame[1:] - t_last_frame[:-1]
intertrial_missed_frames = np.int32(np.round(intertrial_duration * frate)) - 1
# initialize the full times array
frame_times = np.zeros(n_frames + int(np.sum(intertrial_missed_frames)))
ii = 0
for trial, cam_time in enumerate(cam_times):
if cam_time is not None:
# populate first the recovered times within the trials
frame_times[ii: ii + cam_time.size] = cam_time
ii += cam_time.size
if trial == (len(cam_times) - 1):
break
# then extrapolate in-between
nmiss = intertrial_missed_frames[trial]
frame_times[ii: ii + nmiss] = (cam_time[-1] + intertrial_duration[trial] /
(nmiss + 1) * (np.arange(nmiss) + 1))
ii += nmiss
assert all(np.diff(frame_times) > 0) # negative diffs implies a big problem
return frame_times
[docs]def align_with_audio(timestamps, audio, pin_state, count,
extrapolate_missing=True, display=False):
"""
Groom the raw FPGA or Bpod camera timestamps using the frame embedded audio TTLs and frame
counter.
:param timestamps: An array of raw FPGA or Bpod camera timestamps
:param audio: An array of FPGA or Bpod audio TTL times
:param pin_state: An array of camera pin states
:param count: An array of frame numbers
:param extrapolate_missing: If true and the number of timestamps is fewer than the number of
frame counts, the remaining timestamps are extrapolated based on the frame rate, otherwise
they are NaNs
:param display: Plot the resulting timestamps
:return: The corrected frame timestamps
"""
# Some assertions made on the raw data
# assert count.size == pin_state.size, 'frame count and pin state size mismatch'
assert all(np.diff(count) > 0), 'frame count not strictly increasing'
assert all(np.diff(timestamps) > 0), 'FPGA/Bpod camera times not strictly increasing'
same_n_ttl = pin_state['times'].size == audio['times'].size
assert same_n_ttl, 'more audio TTLs detected on camera than TTLs sent'
"""Here we will ensure that the FPGA camera times match the number of video frames in
length. We will make the following assumptions:
1. The number of FPGA camera times is equal to or greater than the number of video frames.
2. No TTLs were missed between the camera and FPGA.
3. No pin states were missed by Bonsai.
4 No pixel count data was missed by Bonsai.
In other words the count and pin state arrays accurately reflect the number of frames
sent by the camera and should therefore be the same length, and the length of the frame
counter should match the number of saved video frames.
The missing frame timestamps are removed in three stages:
1. Remove any timestamps that occurred before video frame acquisition in Bonsai.
2. Remove any timestamps where the frame counter reported missing frames, i.e. remove the
dropped frames which occurred throughout the session.
3. Remove the trailing timestamps at the end of the session if the camera was turned off
in the wrong order.
"""
# Align on first pin state change
first_uptick = pin_state['indices'][0]
first_ttl = np.searchsorted(timestamps, audio['times'][0])
"""Here we find up to which index in the FPGA times we discard by taking the difference
between the index of the first pin state change (when the audio TTL was reported by the
camera) and the index of the first audio TTL in FPGA time. We subtract the difference
between the frame count at the first pin state change and the index to account for any
video frames that were not saved during this period (we will remove those from the
camera FPGA times later).
"""
# Minus any frames that were dropped between the start of frame acquisition and the
# first TTL
start = first_ttl - first_uptick - (count[first_uptick] - first_uptick)
# Get approximate frame rate for extrapolating timestamps (if required)
frate = round(1 / np.nanmedian(np.diff(timestamps)))
if start < 0:
n_missing = abs(start)
_logger.warning(f'{n_missing} missing FPGA/Bpod timestamp(s) at start; '
f'{"extrapolating" if extrapolate_missing else "prepending nans"}')
to_app = (timestamps[0] - (np.arange(n_missing, 0, -1) + 1) / frate
if extrapolate_missing
else np.full(n_missing, np.nan))
timestamps = np.r_[to_app, timestamps] # Prepend the missing times
start = 0
# Remove the extraneous timestamps from the beginning and end
end = count[-1] + 1 + start
ts = timestamps[start:end]
n_missing = count[-1] - ts.size + 1
if n_missing > 0:
# if (n_missing := count[-1] - ts.size + 1) > 0: # py3.8
"""
For ephys sessions there may be fewer FPGA times than frame counts if SpikeGLX is turned
off before the video acquisition workflow. For Bpod this always occurs because Bpod
finishes before the camera workflow. For Bpod the times are already extrapolated for
these late frames."""
_logger.warning(f'{n_missing} fewer FPGA/Bpod timestamps than frame counts; '
f'{"extrapolating" if extrapolate_missing else "appending nans"}')
to_app = ((np.arange(n_missing, ) + 1) / frate + ts[-1]
if extrapolate_missing
else np.full(n_missing, np.nan))
ts = np.r_[ts, to_app] # Append the missing times
assert ts.size >= count.size, 'fewer timestamps than frame counts'
assert ts.size == count[-1] + 1, 'more frames recorded in frame count than timestamps '
# Remove the rest of the dropped frames
ts = ts[count]
assert np.searchsorted(ts, audio['times'][0]) == first_uptick,\
'time of first audio TTL doesn\'t match after alignment'
if ts.size != count.size:
_logger.error('number of timestamps and frames don\'t match after alignment')
if display:
# Plot to check
fig, axes = plt.subplots(1, 1)
y = within_ranges(np.arange(ts.size), pin_state['indices'].reshape(-1, 2)).astype(float)
y *= 1e-5 # For scale when zoomed in
axes.plot(ts, y, marker='d', color='blue', drawstyle='steps-pre', label='GPIO')
axes.plot(ts, np.zeros_like(ts), 'kx', label='FPGA timestamps')
vertical_lines(audio['times'], ymin=0, ymax=1e-5,
color='r', linestyle=':', ax=axes, label='audio TTL')
plt.legend()
return ts
[docs]def attribute_times(arr, events, tol=.1, injective=True, take='first'):
"""
Returns the values of the first array that correspond to those of the second.
Given two arrays of timestamps, the function will return the values of the first array
that most likely correspond to the values of the second. For each of the values in the
second array, the absolute difference is taken and the index of either the first sufficiently
close value, or simply the closest one, is assigned.
If injective is True, once a value has been assigned, to a value it can't be assigned to
another. In other words there is a one-to-one mapping between the two arrays.
:param arr: An array of event times to attribute to those in `events`
:param events: An array of event times considered a subset of `arr`
:param tol: The max absolute difference between values in order to be considered a match
:param injective: If true, once a value has been assigned it will not be assigned again
:param take: If 'first' the first value within tolerance is assigned; if 'nearest' the
closest value is assigned
:returns Numpy array the same length as `values`
"""
take = take.lower()
if take not in ('first', 'nearest'):
raise ValueError('Parameter `take` must be either "first" or "nearest"')
stack = np.ma.masked_invalid(arr, copy=False)
stack.fill_value = np.inf
assigned = np.full(events.shape, -1, dtype=int) # Initialize output array
for i, x in enumerate(events):
dx = np.abs(stack.filled() - x)
if dx.min() < tol: # is any value within tolerance
idx = np.where(dx < tol)[0][0] if take == 'first' else dx.argmin()
assigned[i] = idx
stack.mask[idx] = injective # If one-to-one, remove the assigned value
return assigned
[docs]def groom_pin_state(gpio, audio, ts, tolerance=2., display=False, take='first', min_diff=0.):
"""
Align the GPIO pin state to the FPGA audio TTLs. Any audio TTLs not reflected in the pin
state are removed from the dict and the times of the detected fronts are converted to FPGA
time. At the end of this the number of GPIO fronts should equal the number of audio fronts.
Note:
- This function is ultra safe: we probably don't need assign all the ups and down fronts
separately and could potentially even align the timestamps without removing the missed fronts
- The input gpio and audio dicts may be modified by this function
- For training sessions the frame rate is only 30Hz and the TTLs tend to be broken up by
small gaps. Setting the min_diff to 5ms helps the timestamp assignment accuracy.
:param gpio: array of GPIO pin state values
:param audio: dict of FPGA audio TTLs (see ibllib.io.extractors.ephys_fpga._get_sync_fronts)
:param ts: camera frame times
:param tolerance: two pulses need to be within this many seconds to be considered related
:param take: If 'first' the first value within tolerance is assigned; if 'nearest' the
closest value is assigned
:param display: If true, the resulting timestamps are plotted against the raw audio signal
:param min_diff: Audio TTL fronts less than min_diff seconds apart will be removed
:returns: dict of GPIO FPGA front indices, polarities and FPGA aligned times
:returns: audio times and polarities sans the TTLs not detected in the frame data
:returns: frame times in FPGA time
"""
# Check that the dimensions match
if np.any(gpio['indices'] >= ts.size):
_logger.warning('GPIO events occurring beyond timestamps array length')
keep = gpio['indices'] < ts.size
gpio = {k: gpio[k][keep] for k, v in gpio.items()}
assert audio and audio['times'].size > 0, 'no audio TTLs for session'
assert audio['times'].size == audio['polarities'].size, 'audio data dimension mismatch'
# make sure that there are no 2 consecutive fall or consecutive rise events
assert np.all(np.abs(np.diff(audio['polarities'])) == 2), 'consecutive high/low audio events'
# make sure first TTL is high
assert audio['polarities'][0] == 1
# make sure audio times in order
assert np.all(np.diff(audio['times']) > 0)
# make sure raw timestamps increase
assert np.all(np.diff(ts) > 0), 'timestamps must strictly increase'
# make sure there are state changes
assert gpio['indices'].any(), 'no TTLs detected in GPIO'
# # make sure first GPIO state is high
assert gpio['polarities'][0] == 1
"""
Some audio TTLs appear to be so short that they are not recorded by the camera. These can
be as short as a few microseconds. Applying a cutoff based on framerate was unsuccessful.
Assigning each audio TTL to each pin state change is not easy because some onsets occur very
close together (sometimes < 70ms), on the order of the delay between TTL and frame time.
Also, the two clocks have some degree of drift, so the delay between audio TTL and pin state
change may be zero or even negative.
Here we split the events into audio onsets (lo->hi) and audio offsets (hi->lo). For each
uptick in the GPIO pin state, we take the first audio onset time that was within 100ms of it.
We ensure that each audio TTL is assigned only once, so a TTL that is closer to frame 3 than
frame 1 may still be assigned to frame 1.
"""
ifronts = gpio['indices'] # The pin state flips
audio_times = audio['times']
if ifronts.size != audio['times'].size:
_logger.warning('more audio TTLs than GPIO state changes, assigning timestamps')
to_remove = np.zeros(ifronts.size, dtype=bool) # unassigned GPIO fronts to remove
low2high = ifronts[gpio['polarities'] == 1]
high2low = ifronts[gpio['polarities'] == -1]
assert low2high.size >= high2low.size
# Remove and/or fuse short TTLs
if min_diff > 0:
short, = np.where(np.diff(audio['times']) < min_diff)
audio_times = np.delete(audio['times'], np.r_[short, short + 1])
_logger.debug(f'Removed {short.size * 2} fronts TLLs less than '
f'{min_diff * 1e3:.0f}ms apart')
assert audio_times.size > 0, f'all audio TTLs less than {min_diff}s'
# Onsets
ups = ts[low2high] - ts[low2high][0] # times relative to first GPIO high
onsets = audio_times[::2] - audio_times[0] # audio times relative to first onset
# assign GPIO fronts to audio onset
assigned = attribute_times(onsets, ups, tol=tolerance, take=take)
unassigned = np.setdiff1d(np.arange(onsets.size), assigned[assigned > -1])
if unassigned.size > 0:
_logger.debug(f'{unassigned.size} audio TTL rises were not detected by the camera')
# Check that all pin state upticks could be attributed to an onset TTL
missed = assigned == -1
if np.any(missed):
# if np.any(missed := assigned == -1): # py3.8
_logger.warning(f'{sum(missed)} pin state rises could '
f'not be attributed to an audio TTL')
if display:
ax = plt.subplot()
vertical_lines(ups[assigned > -1],
linestyle='-', color='g', ax=ax,
label='assigned GPIO up state')
vertical_lines(ups[missed],
linestyle='-', color='r', ax=ax,
label='unassigned GPIO up state')
vertical_lines(onsets[unassigned],
linestyle=':', color='k', ax=ax,
alpha=0.3, label='audio onset')
vertical_lines(onsets[assigned],
linestyle=':', color='b', ax=ax, label='assigned audio onset')
plt.legend()
plt.show()
# Remove the missed fronts
to_remove = np.in1d(gpio['indices'], low2high[missed])
assigned = assigned[~missed]
onsets_ = audio_times[::2][assigned]
# Offsets
downs = ts[high2low] - ts[high2low][0]
offsets = audio_times[1::2] - audio_times[1]
assigned = attribute_times(offsets, downs, tol=tolerance, take=take)
unassigned = np.setdiff1d(np.arange(offsets.size), assigned[assigned > -1])
if unassigned.size > 0:
_logger.debug(f'{unassigned.size} audio TTL falls were not detected by the camera')
# Check that all pin state downticks could be attributed to an offset TTL
missed = assigned == -1
if np.any(missed):
# if np.any(missed := assigned == -1): # py3.8
_logger.warning(f'{sum(missed)} pin state falls could '
f'not be attributed to an audio TTL')
# Remove the missed fronts
to_remove |= np.in1d(gpio['indices'], high2low[missed])
assigned = assigned[~missed]
offsets_ = audio_times[1::2][assigned]
# Audio groomed
if np.any(to_remove):
# Check for any orphaned fronts (only one pin state edge was assigned)
to_remove = np.pad(to_remove, (0, to_remove.size % 2), 'edge') # Ensure even size
# Perform xor to find GPIOs where only onset or offset is marked for removal
orphaned = to_remove.reshape(-1, 2).sum(axis=1) == 1
if orphaned.any():
"""If there are orphaned GPIO fronts (i.e. only one edge was assigned to an
audio front), remove the orphaned front its assigned audio TTL. In other words
if both edges cannot be assigned to an audio TTL, we ignore the TTL entirely.
This is a sign that the assignment was bad and extraction may fail."""
_logger.warning('Some onsets but not offsets (or vice versa) were not assigned; '
'this may be a sign of faulty wiring or clock drift')
# Find indices of GPIO upticks where only the downtick was marked for removal
orphaned_onsets, = np.where(~to_remove.reshape(-1, 2)[:, 0] & orphaned)
# The onsets_ array already has the other TTLs removed (same size as to_remove ==
# False) so subtract the number of removed elements from index.
for i, v in enumerate(orphaned_onsets):
orphaned_onsets[i] -= to_remove.reshape(-1, 2)[:v, 0].sum()
# Same for offsets...
orphaned_offsets, = np.where(~to_remove.reshape(-1, 2)[:, 1] & orphaned)
for i, v in enumerate(orphaned_offsets):
orphaned_offsets[i] -= to_remove.reshape(-1, 2)[:v, 1].sum()
# Remove orphaned audio onsets and offsets
onsets_ = np.delete(onsets_, orphaned_onsets[orphaned_onsets < onsets_.size])
offsets_ = np.delete(offsets_, orphaned_offsets[orphaned_offsets < offsets_.size])
_logger.debug(f'{orphaned.sum()} orphaned TTLs removed')
to_remove.reshape(-1, 2)[orphaned] = True
# Remove those unassigned GPIOs
gpio = {k: v[~to_remove[:v.size]] for k, v in gpio.items()}
ifronts = gpio['indices']
# Assert that we've removed discrete TTLs
# A failure means e.g. an up-going front of one TTL was missed
# but not the down-going one.
assert np.all(np.abs(np.diff(gpio['polarities'])) == 2)
assert gpio['polarities'][0] == 1
audio_ = {'times': np.empty(ifronts.size), 'polarities': gpio['polarities']}
audio_['times'][::2] = onsets_
audio_['times'][1::2] = offsets_
else:
audio_ = audio
# Align the frame times to FPGA
fcn_a2b, drift_ppm = dsp.sync_timestamps(ts[ifronts], audio_['times'])
_logger.debug(f'frame audio alignment drift = {drift_ppm:.2f}ppm')
# Add times to GPIO dict
gpio['times'] = fcn_a2b(ts[ifronts])
if display:
# Plot all the onsets and offsets
ax = plt.subplot()
# All Audio TTLS
squares(audio['times'], audio['polarities'],
ax=ax, label='audio TTLs', linestyle=':', color='k', yrange=[0, 1], alpha=0.3)
# GPIO
x = np.insert(gpio['times'], 0, 0)
y = np.arange(x.size) % 2
squares(x, y, ax=ax, label='GPIO')
y = within_ranges(np.arange(ts.size), ifronts.reshape(-1, 2)) # 0 or 1 for each frame
ax.plot(fcn_a2b(ts), y, 'kx', label='cam times')
# Assigned audio
squares(audio_['times'], audio_['polarities'],
ax=ax, label='assigned audio TTL', linestyle=':', color='g', yrange=[0, 1])
ax.legend()
plt.xlabel('FPGA time (s)')
ax.set_yticks([0, 1])
ax.set_title('GPIO - audio TTL alignment')
plt.show()
return gpio, audio_, fcn_a2b(ts)