def _to_dataset(self, data):
""" Converts the data recorded from the BIOSEMI device into a Psychic dataset.
"""
if data == None or data.size == 0:
self.logger.warning(
'Data corrupt: no valid frames found in data packet')
return None
# Undo byte adding that the biosemi has done
data = (data >> 8)
# First channel is status channel
if self.status_as_markers:
Y = (data[:1, :] & 0x00ffff)
else:
Y = numpy.zeros((1, data.shape[1]))
X = data[1:, :] + (2**23) # go from signed to unsigned
print self.reference_channels
if len(self.reference_channels) > 0:
REF = X[self.reference_channels, :]
X = X[self.target_channels, :]
if len(self.reference_channels) > 0:
X = X - numpy.tile(numpy.mean(REF, axis=0), (X.shape[0], 1))
I = self._estimate_timing(X.shape[1])
self.logger.debug('Number of samples parsed: %d' % X.shape[1])
return psychic.DataSet(data=X, labels=Y, ids=I, feat_lab=self.feat_lab)
def _train(self):
""" Train the classifier on a dataset. """
d = self.recorder.read(block=False)
if not d:
raise ClassifierException('First collect some data before training.')
# Save a snapshot of the training data to disk
d.save('test_data.dat')
# Convert markers to classes
Y = np.zeros((3, d.ninstances), dtype=np.bool)
Y[0,:] = (d.Y == 1)[0,:]
Y[1,:] = (d.Y == 2)[0,:]
Y[2,:] = (d.Y == 3)[0,:]
d = psychic.DataSet(labels=Y, cl_lab=['on','off', 'sweep'], default=d)
self._construct_pipeline()
# Train the pipeline
self.classification.train(d)
self.logger.info('Training complete')
# Send a debug plot to client
self._send_debug_image(d)
self.window_node.reset()
self.thres_node.reset()
self.training_complete = True
def _extract_training_trials(self, d):
block_onsets = numpy.flatnonzero(d.Y > 100)
num_blocks = len(block_onsets)
if not num_blocks:
raise ClassifierException(
'No blocks found in recording. Make sure the data is properly labeled.'
)
block_lengths = numpy.hstack(
(numpy.diff(block_onsets), d.ninstances - block_onsets[-1]))
targets = d.Y[0, block_onsets] - 100
options = numpy.unique(targets)
num_options = len(options)
num_instances = num_blocks * num_options
num_channels = d.nfeatures
mdict = {}
for target in sorted(numpy.unique(targets)):
mdict[target] = 'target %02d' % target
# Allocate memory for the blocks
feat_dim_lab = ['channels', 'samples', 'repetitions']
feat_shape = (num_channels,
self.target_window[1] - self.target_window[0],
self.num_repetitions)
X = numpy.zeros(feat_shape + (num_instances, ))
Y = numpy.zeros((2, num_instances))
I = numpy.arange(num_instances)
# Extract each block
for block_num, block_onset, block_length, target in zip(
range(num_blocks), block_onsets, block_lengths, targets):
block = d[block_onset:block_onset + block_length +
self.target_sample_rate]
block = psychic.slice(block, mdict, self.target_window)
#block = psychic.baseline(block, (0, 10))
# Extract each option within a block
for option_num in range(block.nclasses):
if block.get_class(
option_num).ndX.shape[2] < self.num_repetitions:
self.logger.warning(
'could not extract all repetitions of option %d in block %d'
% (option_num + 1, block_num))
continue
instance = block_num * num_options + option_num
X[:, :, :, instance] = block.get_class(
option_num).ndX[:, :, :self.num_repetitions]
Y[0, instance] = (option_num == (target - 1)) # target trial
Y[1, instance] = (option_num !=
(target - 1)) # nontarget trial
# Build new dataset containing the trials
return psychic.DataSet(data=X,
labels=Y,
ids=I,
feat_dim_lab=feat_dim_lab,
l_lab=['target', 'nontarget'])
def _record_data(self):
""" Either generates some random data or extracts a record from the BDF
file. Returns result as a Golem dataset.
"""
self.begin_read_time = self.end_read_time
self.end_read_time = self.begin_read_time + self.buffer_size_seconds
time_to_wait = max(0, self.end_read_time - precision_timer())
time.sleep(time_to_wait)
# Calculate the number of samples to generate
target = int((self.end_read_time - self.T0) * self.sample_rate)
nsamples = target - self.nsamples
if nsamples <= 0:
return None
if self.file_input:
# Determine number of records to read
samples_to_read = nsamples
if self.remaining_frames_in_record != None:
samples_to_read -= self.remaining_frames_in_record.shape[0]
nrecords_to_read = int(
numpy.ceil(
samples_to_read /
float(self.header['record_length'] * self.sample_rate)))
X = self.remaining_frames_in_record
for i in range(nrecords_to_read):
if X == None:
X = self.bdf_reader.read_record()
else:
X = numpy.vstack((X, self.bdf_reader.read_record()))
data_mask = [
i for i, lab in enumerate(self.header['label'])
if lab != 'Status'
]
status_mask = self.header['label'].index('Status')
feat_lab = [self.header['label'][i] for i in data_mask]
self.remaining_frames_in_record = X[nsamples:, :]
Y = X[:nsamples, status_mask].reshape(1, -1).astype(
numpy.int) & 0xffff
X = X[:nsamples, data_mask].T
else:
X = numpy.random.random_integers(self.digital_min,
self.digital_max,
(self.nchannels, nsamples))
feat_lab = self.feat_lab
Y = numpy.zeros((1, nsamples))
I = self._estimate_timing(X.shape[1])
d = psychic.DataSet(data=X, labels=Y, ids=I, feat_lab=feat_lab)
self.nsamples += d.ninstances
return d
def _apply(self, d):
""" Applies classifier to a dataset. """
# Perform preprocessing
d = self.preprocessing.apply(d)
slices = self.slice_node.apply(d)
if slices == None:
return
if self.application_data == None:
self.application_data = slices
else:
self.application_data += slices
repetitions_recorded = numpy.min(self.application_data.ninstances_per_class)
self.logger.info('Repetitions recorded: %d' % repetitions_recorded)
if repetitions_recorded == self.last_repetitions_recorded:
return
self.last_repetitions_recorded = repetitions_recorded
if repetitions_recorded < self.num_repetitions:
return
# Mangle data into shape: (#channels x #samples x #repetitions x #targets)
d = self.application_data
ndX = numpy.zeros(d.feat_shape + (repetitions_recorded, d.nclasses,))
Y = numpy.zeros((2,d.nclasses), dtype=numpy.bool)
I = numpy.arange(d.nclasses)
feat_dim_lab = ['channels', 'samples', 'repetitions']
for cl in range(d.nclasses):
ndX[:,:,:,cl] = d.get_class(cl).ndX[:,:,:repetitions_recorded]
# Update feature labels
feat_lab = list(self.feat_lab)
feat_lab[2] = range(repetitions_recorded)
d = psychic.DataSet(data=ndX, labels=Y, ids=I,
feat_dim_lab=feat_dim_lab,
feat_lab=feat_lab,
cl_lab=['target', 'nontarget'],
default=d)
# Perform actual classification
try:
result = self.classification.apply(d).X
winner = numpy.argmax(result[0,:])
self.logger.info('classification result: %d' % winner)
self.engine.provide_result([list(result[0,:]), winner+1])
self.application_data = None
self.last_repetitions_recorded = 0
self.repetitions_recorded = 0
except ValueError as e:
self.logger.error('Classification failed: %s' % e)
def infer_spatial_pattern(data, y=None, roi_time=None, roi_channels=None,
method='peak'):
"""Estimate the spatial pattern of an ERP component from a psychic DataSet.
This function uses the metadata present in the psychic DataSet object to
provide a convenient interface to the general purpose infer_spatial_pattern
function.
Parameters
----------
data : psychic.DataSet
The trials in the format of a psychic DataSet.
y : 1D array (n_trials,) | 2D array (n_trials, 1) | None
For each trial, a label indicating to which experimental condition
the trial belongs. If None, data.y is used instead. Defaults to None.
roi_time : tuple of floats (start, end) | None
The start and end time (in seconds, end is exclusive) of the time
region of interest. When method='peak', the search for maximum
difference is restricted to this time window. When method='mean', the
mean signal across this time window is used. If None, the entire time
window is used. Defaults to None.
roi_channels : list of floats | None
When method='peak', restrict the search for maximum difference to the
channels with the given indices. When None, do not restrict the search.
Defaults to None.
method : 'peak' | 'mean'
When 'peak', the spatial pattern is the signal at the time of maximum
squared difference between the experimental conditions.
When 'mean', the spatial pattern is the mean difference waveform
between the experimental conditions. Defaults to 'peak'.
Returns
-------
spat_pat : psychic.DataSet
The spatial pattern of the ERP component, stored in a psychic DataSet
object.
"""
if y is None:
y = data.y
channel_names, time = data.feat_lab[:2]
if roi_time is not None:
# Translate between seconds and samples
roi_time = (np.searchsorted(time, roi_time[0]),
np.searchsorted(time, roi_time[1]))
if roi_channels is not None:
# Translate between channel names and indices
roi_channels = [channel_names.index(ch) for ch in roi_channels]
spat_pat = template.infer_spatial_pattern(
data.data.transpose(2, 0, 1), y, roi_time, roi_channels, method)
return psychic.DataSet(spat_pat, ids=channel_names)
def _add_markers(self, d):
""" Label the data with markers. """
self.marker_lock.acquire()
if self.current_marker.type == 'trigger':
Y = numpy.zeros((1, d.ninstances))
else:
Y = numpy.repeat([[self.current_marker.code]],
d.ninstances,
axis=1)
future_markers = []
for m in self.markers:
# Determine the location of the marker in the datastream
y_index = numpy.searchsorted(d.I[0, :], m.timestamp - self.T0)
if y_index <= 0:
# timestamp lies in the past, oh dear!
# mark the first sample, the marker is delayed.
self.current_marker = m
if m.type == 'trigger':
Y[0, 0] = m.code
else:
Y[0, 0:] = m.code
elif y_index >= d.ninstances:
# timestamp lies in the future, save for a later time
future_markers += [m]
else:
# timestamp is present in current data segment
self.current_marker = m
if m.type == 'trigger':
Y[0, y_index] = m.code
else:
Y[0, y_index:] = m.code
# Write some debug info
self.logger.debug('For marker %s, found y_index of %d, (T0=%f)' %
(m, y_index, self.T0))
if y_index < d.ninstances:
self.markerlog.write(
'%f, %f, %d, %d, %f, %f\n' %
(m.timestamp, m.time_received, m.code, y_index,
m.timestamp - self.T0, d.I[0, 0]))
self.markerlog.flush()
self.markers = future_markers
self.marker_lock.release()
return psychic.DataSet(labels=Y, default=d)
def _to_dataset(self, data):
""" Converts the data recorded from the EPOC device into a Psychic dataset.
"""
if data == None or data.size == 0:
self.logger.warning(
'Data corrupt: no valid frames found in data packet')
return None
X = data
Y = numpy.zeros((1, X.shape[1]))
I = self._estimate_timing(X.shape[1])
self.logger.debug('Number of samples parsed: %d' % X.shape[1])
return psychic.DataSet(data=X, label=Y, ids=I, feat_lab=self.feat_lab)
def _to_dataset(self, data):
""" Converts the data recorded from the EPOC device into a Psychic dataset.
"""
if data == None or len(data) == 0:
return None
X = []
for i in range(0, len(data) / self.bytes_per_sample):
packet = data[i * self.bytes_per_sample:(i + 1) *
self.bytes_per_sample]
raw_data = self._cipher.decrypt(
packet[:16]) + self._cipher.decrypt(packet[16:])
X.append([
self._get_level(raw_data, bits)
for bits in sensorBits.values()
])
X = numpy.array(X).T
Y = numpy.zeros((1, X.shape[1]))
I = self._estimate_timing(X.shape[1])
self.logger.debug('Number of samples parsed: %d' % X.shape[1])
return psychic.DataSet(data=X, labels=Y, ids=I, feat_lab=self.feat_lab)
def _to_dataset(self, data):
""" Converts the data recorded from the BIOSEMI device into a Psychic dataset.
"""
if data == None or len(data) == 0:
return None
# Convert data from little-endian 32-bit ints to python integers
data = struct.unpack('<%dI' % (len(data) / 4), data)
# Prepend data recorded earlier that did not cover complete frames
if len(self.unfinished_frames) > 0:
data = self.unfinished_frames + data
self.unfinished_frames = []
# Check sync byte
if data[0] != SYNC_BV:
self.logger.warning('sync lost, trying to find it again')
for i in range(len(data) - 1):
if data[i] == SYNC_BV and data[i + 1] != SYNC_BV:
self.logger.warning('signal re-synced')
data = data[i:]
break
else:
self.logger.warning(
'unable to re-sync signal, discarding data')
return None
# Only keep complete frames
samples_read = len(data) / self.stride
self.unfinished_frames = data[self.stride * samples_read:]
data = data[:self.stride * samples_read]
frames = numpy.array(data,
dtype=numpy.uint32).reshape(-1, self.stride).T
# Test if sync markers line up,,
if len(numpy.flatnonzero(frames[0, :] != SYNC_BV)) > 0:
self.logger.warning('sync lost, discarding data')
return None
self.battery_low = len(
numpy.flatnonzero(frames[1, :] & BATTERY_BV)) > 0
self.cms_in_range = len(
numpy.flatnonzero(frames[1, :] & CMS_RANGE_BV) == 0) > 0
# Keep only first 32 channels + 8 external ones
frames = frames[range(1, 33) + range(257, 266), :]
# Undo byte adding that the biosemi has done
frames = (frames >> 8)
# First channel is status channel
if self.status_as_markers:
Y = (frames[:1, :] & 0x00ffff)
else:
Y = numpy.zeros((1, frames.shape[1]))
X = frames[1:, :] + (2**23) # go from signed to unsigned
# Calculate reference signal
if len(self.reference_channels) > 0:
REF = X[self.reference_channels, :]
X = X[self.target_channels, :]
# Re-reference the signal to the chosen reference
if len(self.reference_channels) > 0:
X = X - numpy.tile(numpy.mean(REF, axis=0), (X.shape[0], 1))
I = self._estimate_timing(X.shape[1])
self.logger.debug('Number of samples parsed: %d' % X.shape[1])
return psychic.DataSet(data=X, labels=Y, ids=I, feat_lab=self.feat_lab)
def _raw_to_dataset(self, data):
""" Decodes a string of raw data read from the IMEC device into a Psychic
dataset """
num_bytes = len(data)
self.logger.debug('Handling datapacket of length %d' % num_bytes)
if num_bytes < self.bytes_per_frame:
self.logger.warning('Data incomplete: read at least %d bytes before'
' calling this function' % self.bytes_per_frame)
return [None, bytes('')]
#data = struct.unpack('%dB' % num_bytes, data_string)
# Construct dataset from the raw data
samples = []
first_frame = True
i = 0
while i <= num_bytes-self.bytes_per_frame:
# Search for the next frame. This frame should be right next to the
# frame we just parsed. But sometimes, the device inserts some
# bogus values between frames, which we need to skip
frame_found = False
frame_index = i
for j in range(i, num_bytes-self.bytes_per_frame+1):
# Data should begin with sync byte ('S' == 0x53)
if data[j] != 0x53:
continue
# Next frame should also begin with sync byte
if (j < num_bytes-2*self.bytes_per_frame and
data[j+self.bytes_per_frame] != 0x53):
continue
# Battery level should be between 120 and 165
if data[j+2] < 120 or data[j+2] > 165:
continue
frame_found = True
frame_index = j
break
if frame_index - i > 0:
self.logger.debug('garbage bytes: %d' % (frame_index - i))
i = frame_index
if not frame_found:
# Done with this data packet
break
if first_frame:
self.logger.debug('First frame found on index %d, seq number '
'%d' % (i, data[i+1]))
first_frame = False
frame = self._decode_frame(data[i:i+self.bytes_per_frame])
# Determine number of dropped frames. Note that if more than 255
# frames are dropped, this does not work.
if self.last_frame == None:
dropped_frames = 0
elif frame.seq > self.last_frame.seq:
dropped_frames = (frame.seq-1) - self.last_frame.seq
elif frame.seq < self.last_frame.seq:
dropped_frames = (frame.seq+255) - self.last_frame.seq
else:
self.logger.warning('Data corrupt: duplicate frame number in '
'data packet (%d = %d), i was %d' %
(self.last_frame.seq, frame.seq, i))
# don't use this frame
i += self.bytes_per_frame
# fix dropped frames when we restore accurate sequence numbers
dropped_frames = 0
continue
if dropped_frames > 0:
self.logger.warning('Dropped %d frames' % dropped_frames)
self.droppedframeslog.write('%f, %f, %d\n' %
(precision_timer(), self.last_id, dropped_frames))
self.droppedframeslog.flush()
# Interpolate the dropped frames if possible
for j in range(1, dropped_frames+1):
if self.last_frame != None:
inter = ( self.last_frame.X[:,1] +
j * (frame.X[:,0]-self.last_frame.X[:,1]) /
float(dropped_frames+1) )
samples.append(numpy.vstack((inter, inter)).T)
else:
samples.append(numpy.vstack((frame.X[:,0], frame.X[:,0])).T)
# Append the current frame
samples.append(frame.X)
self.last_frame = frame
i += self.bytes_per_frame
if len(samples) == 0:
return (None, data)
X = numpy.hstack(samples)[self.target_channels,:]
Y = numpy.zeros([1, X.shape[1]])
I = self._estimate_timing(X.shape[1])
self.logger.debug('Number of bytes parsed: %d' % i)
d = psychic.DataSet(data=X, labels=Y, ids=I, feat_lab=self.feat_lab)
return (d, data[i:])
def _generate_debug_image(self, d):
""" Generate image describing the training data. """
d = psychic.DataSet(cl_lab=self.cl_lab, default=d)
fig = psychic.plot_erp(d, enforce_equal_n=False)
fig.set_size_inches(7, 11)
return fig
def infer_temporal_pattern(data, spat_bf, y=None, refine=None,
refine_params=None):
"""Estimate temporal pattern of an ERP component from a psychic DataSet.
This function uses the metadata present in the psychic DataSet object to
provide a convenient interface to the general purpose
infer_temporal_pattern function.
Parameters
----------
data : psychic.DataSet
The trials in the format of a psychic DataSet.
spat_bf : instance of LCMV
The spatial beamformer that will be used to extract the ERP timecourse.
y : 1D array (n_trials,) | 2D array (n_trials, 1) | None
For each trial, a label indicating to which experimental condition
the trial belongs. If None, data.y is used instead. Defaults to None.
refine : 'zero' | 'peak-mean' | 'thres' | 'guass' | None
The method used to refine the template:
'zero': Zero out everything outside the time region of interest.
'peak-mean': Find the peak inside the time region of interest. Then,
find the points before and after the peak, where the
signal drops below the average signal outside the time
region of interest. Zero out everything outside those
points.
'thres': As well as zero-ing out everything outside the time
region of interest, also zero out any part of the signal
which amplitude is below 4 standard deviations of the
signal amplitude during the baseline period.
'gauss': Multiply the signal with a Gaussian kernel that is defined
over time.
Defaults to None, which means no refining of the template is performed.
refine_params : dict | None
Parameters for the chosen refining method. Each method uses different
parameters taken from this dictionary. Possible parameters are:
Used by 'zero', 'peak-mean' and 'thres':
roi_time : tuple of floats
The start and end time (in seconds, end is exclusive) of the time
region of interest.
Used by 'gauss':
mu : float
Time at which to center the Gaussian kernel in seconds.
sigma : float
Standard deviation (in seconds) of the Gaussian kernel.
Returns
-------
temp_pat : psychic.DataSet
The temporal pattern of the ERP component, stored in a psychic DataSet
object.
"""
if y is None:
y = data.y
channel_names, time = data.feat_lab[:2]
if refine_params is None:
refine_params = dict()
# Translate between seconds and samples
if 'roi_time' in refine_params:
roi_time = refine_params['roi_time']
roi_time = (np.searchsorted(time, roi_time[0]),
np.searchsorted(time, roi_time[1]))
refine_params['roi_time'] = roi_time
if 'mu' in refine_params:
refine_params['mu'] = np.searchsorted(time, refine_params['mu'])
if 'sigma' in refine_params:
refine_params['sigma'] = np.searchsorted(time, refine_params['sigma'])
baseline_time = (0, np.searchsorted(time, 0))
temp_pat = template.infer_temporal_pattern(
data.data.transpose(2, 0, 1), y, spat_bf, baseline_time, refine,
refine_params)
return psychic.DataSet(temp_pat, ids=time)
def run(self):
""" Don't call this directly. Use start() and start_capture() to start
reading data from the device. """
self.running = True
# We keep track of the last 10 seconds of estimations of the sample rate
# of the device in a circular buffer.
self.estimated_sample_rates = collections.deque(
maxlen=numpy.ceil(10 / float(self.buffer_size_seconds)))
# We keep track of a 'drift table'. The first row contains the
# timestamps of the incomining data. The second row contains the number
# of samples read so far. The true samplerate of the device can be
# estimated by performing a linear regression on the number of samples
# read versus the timestamps.
npoints = numpy.ceil(60 / float(self.buffer_size_seconds))
self._drift_table = [
collections.deque(maxlen=npoints),
collections.deque(maxlen=npoints)
]
# Open BDF file output
if self.bdf_file != None:
self.bdf_writer = psychic.BDFWriter(self.bdf_file,
self.sample_rate,
self.nchannels)
self._set_bdf_values()
self.bdf_writer.write_header()
self.file_output = True
else:
self.file_output = False
# Perform initialization of the device driver, which returns the
# timestamp of the first data packet
self.T0 = self._open()
self.last_id = 0
while self.running:
try:
# Record some data
d = self._record_data()
# Check whether the decoding of the data succeeded
if d == None:
continue
# Add markers to the data
d = self._add_markers(d)
# Write data without gain factor to file
if self.file_output:
self.bdf_writer.write_raw(d)
# Check whether calibration period is complete
if precision_timer() > self.T0 + self.calibration_time:
self.calibrated_event.set()
if self.capture_data:
# Apply gain factor to data, producing values that
# correspond to actual voltage
d = psychic.DataSet(d.data * self.gain + self.physical_min,
default=d)
# Append the data to the buffer and notify
# any listeners (usually the classifier)
self.data_condition.acquire()
if self.data == None:
self.data = d
else:
self.data += d
self.data_condition.notify()
self.data_condition.release()
except IOError, e:
self.logger.error('I/O Error: %s' % e)
raise