def _cochlear_trim_sai_marginals(filename_and_indexes):
try:
filename, norm_segstart, norm_segend, audio_id, NAP_detail = filename_and_indexes
sai_video_filename = '{}_sai_video_{}'.format(filename, NAP_detail)
if os.path.isfile('{}.npy'.format(sai_video_filename)):
return sai_video_filename
if NAP_detail == 'high':
try:
NAP = utils.csv_to_array(filename + 'cochlear' + NAP_detail)
except:
NAP = brain.cochlear(filename,
stride=1,
rate=44100,
apply_filter=0,
suffix='cochlear' + NAP_detail)
if NAP_detail == 'low':
try:
NAP = utils.csv_to_array(filename + 'cochlear' + NAP_detail)
except:
NAP = brain.cochlear(
filename,
stride=IO.NAP_STRIDE,
rate=IO.NAP_RATE,
apply_filter=0,
suffix='cochlear' + NAP_detail
) # Seems to work best, in particular when they are all the same.
num_channels = NAP.shape[1]
input_segment_width = 2048
sai_params = CreateSAIParams(num_channels=num_channels,
input_segment_width=input_segment_width,
trigger_window_width=input_segment_width,
sai_width=1024)
sai = pysai.SAI(sai_params)
NAP = utils.trim_right(NAP[np.int(np.rint(NAP.shape[0] *
norm_segstart)):np.
int(np.rint(NAP.shape[0] * norm_segend))],
threshold=.05)
sai_video = [
np.copy(sai.RunSegment(input_segment.T))
for input_segment in utils.chunks(NAP, input_segment_width)
]
del NAP
np.save(sai_video_filename,
np.array([sai_rectangles(frame) for frame in sai_video]))
return sai_video_filename
except:
print utils.print_exception(
'Calculation SAI video failed for file {}, NAP detail {}'.format(
filename, NAP_detail))
return False
def testInputSegmentWidthIsLargerThanBuffer(self):
params = CreateSAIParams(num_channels=2,
sai_width=10,
input_segment_width=200,
trigger_window_width=200)
sai = pysai.SAI(params)
params.trigger_window_width = params.input_segment_width // 10
self.assertGreater(
params.input_segment_width,
params.num_triggers_per_frame * params.trigger_window_width)
self.assertRaises(AssertionError, sai.Redesign, params)
def testInputSegmentWidthSmallerThanTriggerWindow(self):
"""Tests small hop between segments."""
num_channels = 1
total_input_samples = 20
period = 5
full_input = CreatePulseTrain(num_channels, total_input_samples,
period)
num_frames = 4
input_segment_width = total_input_samples // num_frames
sai_params = CreateSAIParams(num_channels=num_channels,
input_segment_width=input_segment_width,
trigger_window_width=total_input_samples,
sai_width=15)
self.assertLess(sai_params.input_segment_width,
sai_params.trigger_window_width)
sai_params.future_lags = 0 # Only compute past lags.
self.assertGreaterEqual(period, input_segment_width)
sai = pysai.SAI(sai_params)
for i in range(num_frames):
segment = (full_input[:, i * input_segment_width:(i + 1) *
input_segment_width])
sai_frame = sai.RunSegment(segment)
print("Frame {}\nInput\n{}\nOutput\n{}".format(
i, segment, sai_frame))
self.assertNotEqual(np.abs(segment).sum(), 0)
# Since the input segment is never all zero, there should always
# be a peak at zero lag.
sai_channel = sai_frame[0, :]
self.assertTrue(HasPeakAt(sai_channel, len(sai_channel) - 1))
if i == 0:
# Since the pulse train period is larger than the input segment
# size, the first input segment will only see a single impulse,
# most of the SAI will be zero.
np.testing.assert_allclose(sai_channel[:len(sai_channel) - 1],
np.zeros(len(sai_channel) - 1),
1e-9)
if i == num_frames - 1:
# By the last frame, the SAI's internal buffer will have
# accumulated the full input signal, so the resulting image
# should contain kPeriod peaks.
for j in range(len(sai_channel) - 1, 0, -period):
self.assertTrue(HasPeakAt(sai_channel, j))
def testInputWidthDoesntMatchInputSegmentWidth(self):
num_channels = 2
input_segment_width = 10
segment = CreatePulseTrain(num_channels, input_segment_width, period=4)
sai_width = 20
expected_input_segment_width = input_segment_width - 1
sai_params = CreateSAIParams(
num_channels=num_channels,
sai_width=sai_width,
input_segment_width=expected_input_segment_width,
trigger_window_width=sai_width + 1)
self.assertNotEqual(sai_params.input_segment_width,
input_segment_width)
sai = pysai.SAI(sai_params)
self.assertRaises(AssertionError, sai.RunSegment, segment)
def testMatchesMatlabOnBinauralData(self):
test_name = "binaural_test"
input_segment_width = 882
num_channels = 71
# The Matlab CARFAC output is transposed compared to the C++.
input_segment = LoadMatrix(test_name + "-matlab-nap1.txt",
input_segment_width,
num_channels).transpose()
sai_params = CreateSAIParams(num_channels=num_channels,
input_segment_width=input_segment_width,
trigger_window_width=input_segment_width,
sai_width=500)
sai = pysai.SAI(sai_params)
sai_frame = sai.RunSegment(input_segment)
expected_sai_frame = LoadMatrix(test_name + "-matlab-sai1.txt",
num_channels, sai_params.sai_width)
np.testing.assert_allclose(expected_sai_frame, sai_frame, rtol=1e-5)
WriteMatrix(test_name + "-py-sai1.txt", sai_frame)
def _RunMultiChannelPulseTrainTest(self, period, num_channels):
input_segment_width = 38
segment = CreatePulseTrain(num_channels, input_segment_width, period)
sai_params = CreateSAIParams(num_channels=num_channels,
input_segment_width=input_segment_width,
trigger_window_width=input_segment_width,
sai_width=15)
sai_params.future_lags = 0 # Only compute past lags.
sai = pysai.SAI(sai_params)
sai_frame = sai.RunSegment(segment)
# The output should have peaks at the same positions, regardless of
# input phase.
for i in range(num_channels):
sai_channel = sai_frame[i, :]
for j in range(len(sai_channel) - 1, 0, -period):
print(i, j, sai_channel, HasPeakAt(sai_channel, j))
self.assertTrue(HasPeakAt(sai_channel, j))
print("Input\n{}\nOutput\n{}".format(segment, sai_frame))
