def test_find_bad_by_ransac(raw_tmp):
"""Test the RANSAC component of NoisyChannels."""
# Set a consistent random seed for all RANSAC runs
RANSAC_RNG = 435656
# RANSAC identifies channels that go bad together and are highly correlated.
# Inserting highly correlated signal in channels 0 through 6 at 30 Hz
raw_tmp._data[0:6, :] = _generate_signal(30, 30, raw_tmp.times)
# Run different variations of RANSAC on the same data
test_matrix = {
# List items represent [matlab_strict, channel_wise, max_chunk_size]
"by_window": [False, False, None],
"by_channel": [False, True, None],
"by_channel_maxchunk": [False, True, 2],
"by_window_strict": [True, False, None],
"by_channel_strict": [True, True, None],
}
bads = {}
corr = {}
for name, args in test_matrix.items():
nd = NoisyChannels(raw_tmp,
do_detrend=False,
random_state=RANSAC_RNG,
matlab_strict=args[0])
nd.find_bad_by_ransac(channel_wise=args[1], max_chunk_size=args[2])
# Save bad channels and RANSAC correlation matrix for later comparison
bads[name] = nd.bad_by_ransac
corr[name] = nd._extra_info["bad_by_ransac"]["ransac_correlations"]
# Test whether all methods detected bad channels properly
assert bads["by_window"] == raw_tmp.ch_names[0:6]
assert bads["by_channel"] == raw_tmp.ch_names[0:6]
assert bads["by_channel_maxchunk"] == raw_tmp.ch_names[0:6]
assert bads["by_window_strict"] == raw_tmp.ch_names[0:6]
assert bads["by_channel_strict"] == raw_tmp.ch_names[0:6]
# Make sure non-strict correlation matrices all match
assert np.allclose(corr["by_window"], corr["by_channel"])
assert np.allclose(corr["by_window"], corr["by_channel_maxchunk"])
# Make sure MATLAB-strict correlation matrices match
assert np.allclose(corr["by_window_strict"], corr["by_channel_strict"])
# Make sure strict and non-strict matrices differ
assert not np.allclose(corr["by_window"], corr["by_window_strict"])
# Ensure that RANSAC doesn't change random state if in MATLAB-strict mode
rng = RandomState(RANSAC_RNG)
init_state = rng.get_state()[2]
nd = NoisyChannels(raw_tmp,
do_detrend=False,
random_state=rng,
matlab_strict=True)
nd.find_bad_by_ransac()
assert rng.get_state()[2] == init_state
def test_find_bad_by_ransac_err(raw_tmp):
"""Test error handling in the `find_bad_by_ransac` method."""
# Set n_samples very very high to trigger a memory error
n_samples = int(1e100)
nd = NoisyChannels(raw_tmp, do_detrend=False)
with pytest.raises(MemoryError):
nd.find_bad_by_ransac(n_samples=n_samples)
# Set n_samples to a float to trigger a type error
n_samples = 35.5
nd = NoisyChannels(raw_tmp, do_detrend=False)
with pytest.raises(TypeError):
nd.find_bad_by_ransac(n_samples=n_samples)
# Test IOError when too few good channels for RANSAC sample size
n_chans = raw_tmp._data.shape[0]
nd = NoisyChannels(raw_tmp, do_detrend=False)
nd.bad_by_deviation = raw_tmp.info["ch_names"][0:int(n_chans * 0.8)]
with pytest.raises(IOError):
nd.find_bad_by_ransac()
# Test IOError when not enough channels for RANSAC predictions
raw_tmp._data[0:(n_chans - 2), :] = 0 # make all channels flat except 2
nd = NoisyChannels(raw_tmp, do_detrend=False)
with pytest.raises(IOError):
nd.find_bad_by_ransac()
raw = mne.io.RawArray(X, info)
raw.set_montage(montage, verbose=False)
###############################################################################
# Assign the mne object to the :class:`NoisyChannels` class. The resulting object
# will be the place where all following methods are performed.
nd = NoisyChannels(raw, random_state=1337)
nd2 = NoisyChannels(raw, random_state=1337)
###############################################################################
# Find all bad channels using channel-wise RANSAC and print a summary
start_time = perf_counter()
nd.find_bad_by_ransac(channel_wise=True)
print("--- %s seconds ---" % (perf_counter() - start_time))
# Repeat channel-wise RANSAC using a single channel at a time. This is slower
# but needs less memory.
start_time = perf_counter()
nd2.find_bad_by_ransac(channel_wise=True, max_chunk_size=1)
print("--- %s seconds ---" % (perf_counter() - start_time))
###############################################################################
# Now the bad channels are saved in `bads` and we can continue processing our
# `raw` object. For more information, we can access attributes of the ``nd``
# instance:
# Check channels that go bad together by correlation (RANSAC)
print(nd.bad_by_ransac)
def test_findnoisychannels(raw, montage):
raw.set_montage(montage)
nd = NoisyChannels(raw)
nd.find_all_bads(ransac=True)
bads = nd.get_bads()
iterations = (
10 # remove any noisy channels by interpolating the bads for 10 iterations
)
for iter in range(0, iterations):
raw.info["bads"] = bads
raw.interpolate_bads()
nd = NoisyChannels(raw)
nd.find_all_bads(ransac=True)
bads = nd.get_bads()
# make sure no bad channels exist in the data
raw.drop_channels(ch_names=bads)
# Test for NaN and flat channels
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
# Insert a nan value for a random channel
rand_chn_idx1 = int(np.random.randint(0, m, 1))
rand_chn_idx2 = int(np.random.randint(0, m, 1))
rand_chn_lab1 = raw_tmp.ch_names[rand_chn_idx1]
rand_chn_lab2 = raw_tmp.ch_names[rand_chn_idx2]
raw_tmp._data[rand_chn_idx1, n - 1] = np.nan
raw_tmp._data[rand_chn_idx2, :] = np.ones(n)
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_nan_flat()
assert nd.bad_by_nan == [rand_chn_lab1]
assert nd.bad_by_flat == [rand_chn_lab2]
# Test for high and low deviations in EEG data
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
# Now insert one random channel with very low deviations
rand_chn_idx = int(np.random.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
raw_tmp._data[rand_chn_idx, :] = raw_tmp._data[rand_chn_idx, :] / 10
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_deviation()
assert rand_chn_lab in nd.bad_by_deviation
# Inserting one random channel with a high deviation
raw_tmp = raw.copy()
rand_chn_idx = int(np.random.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
arbitrary_scaling = 5
raw_tmp._data[rand_chn_idx, :] *= arbitrary_scaling
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_deviation()
assert rand_chn_lab in nd.bad_by_deviation
# Test for correlation between EEG channels
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(np.random.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# Use cosine instead of sine to create a signal
low = 10
high = 30
n_freq = 5
signal = np.zeros((1, n))
for freq_i in range(n_freq):
freq = np.random.randint(low, high, n)
signal[0, :] += np.cos(2 * np.pi * raw.times * freq)
raw_tmp._data[rand_chn_idx, :] = signal * 1e-6
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_correlation()
assert rand_chn_lab in nd.bad_by_correlation
# Test for high freq noise detection
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(np.random.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# Use freqs between 90 and 100 Hz to insert hf noise
signal = np.zeros((1, n))
for freq_i in range(n_freq):
freq = np.random.randint(90, 100, n)
signal[0, :] += np.sin(2 * np.pi * raw.times * freq)
raw_tmp._data[rand_chn_idx, :] = signal * 1e-6
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_hfnoise()
assert rand_chn_lab in nd.bad_by_hf_noise
# Test for signal to noise ratio in EEG data
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(np.random.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# inserting an uncorrelated high frequency (90 Hz) signal in one channel
raw_tmp[rand_chn_idx, :] = np.sin(2 * np.pi * raw.times * 90) * 1e-6
nd = NoisyChannels(raw_tmp)
nd.find_bad_by_SNR()
assert rand_chn_lab in nd.bad_by_SNR
# Test for finding bad channels by RANSAC
raw_tmp = raw.copy()
# Ransac identifies channels that go bad together and are highly correlated.
# Inserting highly correlated signal in channels 0 through 3 at 30 Hz
raw_tmp._data[0:6, :] = np.cos(2 * np.pi * raw.times * 30) * 1e-6
nd = NoisyChannels(raw_tmp)
np.random.seed(30)
nd.find_bad_by_ransac()
bads = nd.bad_by_ransac
assert bads == raw_tmp.ch_names[0:6]
def test_findnoisychannels(raw, montage):
"""Test find noisy channels."""
# Set a random state for the test
rng = np.random.RandomState(30)
raw.set_montage(montage)
nd = NoisyChannels(raw, random_state=rng)
nd.find_all_bads(ransac=True)
bads = nd.get_bads()
iterations = (
10 # remove any noisy channels by interpolating the bads for 10 iterations
)
for iter in range(0, iterations):
if len(bads) == 0:
continue
raw.info["bads"] = bads
raw.interpolate_bads()
nd = NoisyChannels(raw, random_state=rng)
nd.find_all_bads(ransac=True)
bads = nd.get_bads()
# make sure no bad channels exist in the data
raw.drop_channels(ch_names=bads)
# Test for NaN and flat channels
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
# Insert a nan value for a random channel and make another random channel
# completely flat (ones)
idxs = rng.choice(np.arange(m), size=2, replace=False)
rand_chn_idx1 = idxs[0]
rand_chn_idx2 = idxs[1]
rand_chn_lab1 = raw_tmp.ch_names[rand_chn_idx1]
rand_chn_lab2 = raw_tmp.ch_names[rand_chn_idx2]
raw_tmp._data[rand_chn_idx1, n - 1] = np.nan
raw_tmp._data[rand_chn_idx2, :] = np.ones(n)
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_nan_flat()
assert nd.bad_by_nan == [rand_chn_lab1]
assert nd.bad_by_flat == [rand_chn_lab2]
# Test for high and low deviations in EEG data
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
# Now insert one random channel with very low deviations
rand_chn_idx = int(rng.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
raw_tmp._data[rand_chn_idx, :] = raw_tmp._data[rand_chn_idx, :] / 10
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_deviation()
assert rand_chn_lab in nd.bad_by_deviation
# Inserting one random channel with a high deviation
raw_tmp = raw.copy()
rand_chn_idx = int(rng.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
arbitrary_scaling = 5
raw_tmp._data[rand_chn_idx, :] *= arbitrary_scaling
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_deviation()
assert rand_chn_lab in nd.bad_by_deviation
# Test for correlation between EEG channels
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(rng.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# Use cosine instead of sine to create a signal
low = 10
high = 30
n_freq = 5
signal = np.zeros((1, n))
for freq_i in range(n_freq):
freq = rng.randint(low, high, n)
signal[0, :] += np.cos(2 * np.pi * raw.times * freq)
raw_tmp._data[rand_chn_idx, :] = signal * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_correlation()
assert rand_chn_lab in nd.bad_by_correlation
# Test for high freq noise detection
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(rng.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# Use freqs between 90 and 100 Hz to insert hf noise
signal = np.zeros((1, n))
for freq_i in range(n_freq):
freq = rng.randint(90, 100, n)
signal[0, :] += np.sin(2 * np.pi * raw.times * freq)
raw_tmp._data[rand_chn_idx, :] = signal * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_hfnoise()
assert rand_chn_lab in nd.bad_by_hf_noise
# Test for signal to noise ratio in EEG data
raw_tmp = raw.copy()
m, n = raw_tmp._data.shape
rand_chn_idx = int(rng.randint(0, m, 1))
rand_chn_lab = raw_tmp.ch_names[rand_chn_idx]
# inserting an uncorrelated high frequency (90 Hz) signal in one channel
raw_tmp[rand_chn_idx, :] = np.sin(2 * np.pi * raw.times * 90) * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_SNR()
assert rand_chn_lab in nd.bad_by_SNR
# Test for finding bad channels by RANSAC
raw_tmp = raw.copy()
# Ransac identifies channels that go bad together and are highly correlated.
# Inserting highly correlated signal in channels 0 through 3 at 30 Hz
raw_tmp._data[0:6, :] = np.cos(2 * np.pi * raw.times * 30) * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_ransac()
bads = nd.bad_by_ransac
assert bads == raw_tmp.ch_names[0:6]
# Test for finding bad channels by channel-wise RANSAC
raw_tmp = raw.copy()
# Ransac identifies channels that go bad together and are highly correlated.
# Inserting highly correlated signal in channels 0 through 3 at 30 Hz
raw_tmp._data[0:6, :] = np.cos(2 * np.pi * raw.times * 30) * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_bad_by_ransac(channel_wise=True)
bads = nd.bad_by_ransac
assert bads == raw_tmp.ch_names[0:6]
# Test not-enough-memory and n_samples type exceptions
raw_tmp = raw.copy()
raw_tmp._data[0:6, :] = np.cos(2 * np.pi * raw.times * 30) * 1e-6
nd = NoisyChannels(raw_tmp, random_state=rng)
# Set n_samples very very high to trigger a memory error
n_samples = int(1e100)
with pytest.raises(MemoryError):
nd.find_bad_by_ransac(n_samples=n_samples)
# Set n_samples to a float to trigger a type error
n_samples = 35.5
with pytest.raises(TypeError):
nd.find_bad_by_ransac(n_samples=n_samples)
# Test IOError when not enough channels for ransac predictions
raw_tmp = raw.copy()
# Make flat all channels except 2
num_bad_channels = raw._data.shape[0] - 2
raw_tmp._data[0:num_bad_channels, :] = np.zeros_like(
raw_tmp._data[0:num_bad_channels, :]
)
nd = NoisyChannels(raw_tmp, random_state=rng)
nd.find_all_bads(ransac=False)
with pytest.raises(IOError):
nd.find_bad_by_ransac()
datatype='eeg',
suffix='eeg')
raw = read_raw_bids(bids_path=bids_path, extra_params=dict(preload=True))
raw.info['bads'] = [
'FT7', 'FT8', 'T7', 'T8', 'TP7', 'TP8', 'AF7', 'AF8', 'M1', 'M2',
'BIP1', 'BIP2', 'BIP3', 'BIP4', 'BIP5', 'BIP6', 'BIP7', 'BIP8', 'BIP9',
'BIP10', 'BIP11', 'BIP12', 'BIP13', 'BIP14', 'BIP15', 'BIP16', 'BIP17',
'BIP18', 'BIP19', 'BIP20', 'BIP21', 'BIP22', 'BIP23', 'BIP24'
]
raw.drop_channels(ch_names=raw.info['bads'])
montage = mne.channels.make_standard_montage('standard_1020')
raw.set_montage(montage)
filt_raw = raw.copy()
filt_raw.drop_channels('EOG')
nd = NoisyChannels(filt_raw)
nd.find_bad_by_ransac(n_samples=50, channel_wise=True)
raw.info['bads'] = nd.bad_by_ransac
raw.notch_filter(np.arange(50, 250, 50))
filt_raw.notch_filter(np.arange(50, 250, 50))
raw.filter(l_freq=settings.high_filter,
h_freq=settings.low_filter,
fir_design='firwin')
filt_raw.filter(l_freq=1, h_freq=settings.low_filter)
filt_raw.info['bads'] = raw.info['bads']
raw.interpolate_bads(reset_bads=True, mode='accurate')
filt_raw.interpolate_bads(reset_bads=True, mode='accurate')
raw.resample(200, npad='auto')
filt_raw.resample(200, npad='auto')
raw.set_eeg_reference('average')
filt_raw.set_eeg_reference('average')
eog_evoked = create_eog_epochs(raw).average()
raw = mne.io.RawArray(X, info)
raw.set_montage(montage, verbose=False)
###############################################################################
# Assign the mne object to the :class:`NoisyChannels` class. The resulting object
# will be the place where all following methods are performed.
nd = NoisyChannels(raw)
nd2 = NoisyChannels(raw)
###############################################################################
# Find all bad channels and print a summary
start_time = perf_counter()
nd.find_bad_by_ransac()
print("--- %s seconds ---" % (perf_counter() - start_time))
# Repeat RANSAC in a channel wise manner. This is slower but needs less memory.
start_time = perf_counter()
nd2.find_bad_by_ransac(channel_wise=True)
print("--- %s seconds ---" % (perf_counter() - start_time))
###############################################################################
# Now the bad channels are saved in `bads` and we can continue processing our
# `raw` object. For more information, we can access attributes of the ``nd``
# instance:
# Check channels that go bad together by correlation (RANSAC)
print(nd.bad_by_ransac)
assert set(bad_ch_names) == set(nd.bad_by_ransac)
freq_good * time)
for i in range(n_chans)
]
# Scale the signal amplitude and add noise.
X = 2e-5 * np.array(X) + 1e-5 * np.random.random((n_chans, time.shape[0]))
raw = mne.io.RawArray(X, info)
raw.set_montage(montage, verbose=False)
###############################################################################
# Assign the mne object to the :class:`NoisyChannels` class. The resulting object
# will be the place where all following methods are performed.
nd = NoisyChannels(raw)
###############################################################################
# Find all bad channels and print a summary
start_time = perf_counter()
nd.find_bad_by_ransac()
print("--- %s seconds ---" % (perf_counter() - start_time))
###############################################################################
# Now the bad channels are saved in `bads` and we can continue processing our
# `raw` object. For more information, we can access attributes of the ``nd``
# instance:
# Check channels that go bad together by correlation (RANSAC)
print(nd.bad_by_ransac)
assert set(bad_ch_names) == set(nd.bad_by_ransac)
