def get_freqs(sample_rate, window_length=0.060, increment=None):
if increment is None:
increment = 2.0 / sample_rate
nt = int(window_length*2*sample_rate)
s = np.random.randn(nt)
pfreq,psd1,ps_var,phase = power_spectrum_jn(s, sample_rate, window_length, increment)
return pfreq
def compute_filt_power_spectra(filts, sample_rate, lags_ms):
filt_freq = None
psds = list()
# fi = (lags_ms > 8.) & (lags_ms < 125.)
# fi = (lags_ms > 8.)
fi = (lags_ms > -1)
for f in filts:
filt_freq,filt_ps,filt_ps_var,filt_phase = power_spectrum_jn(f[fi], sample_rate, 0.250, 0.025)
psds.append(filt_ps)
psds = np.array(psds)
psds /= psds.max()
# log_transform(psds)
fi = filt_freq < 40.
return filt_freq[fi], psds[:, fi]
def get_env_psds(rp):
# compute the stim amp envelope
stim_env = rp.U.sum(axis=1)
stim_env /= stim_env.max()
# grab the amplitude envelope for each stimulus
print 'Computing psds...'
stim_ids = rp.event_df.stim_id.unique()
env_data = {'stim_id': list(), 'stim_type': list(), 'duration': list(), 'xindex': list()}
env_psds = list()
env_freq = None
duration_thresh = 1.1
for stim_id in stim_ids:
i = rp.event_df.stim_id == stim_id
stim_type = rp.event_df[i].stim_type.values[0]
stime = rp.event_df[i].start_time.values[0]
etime = rp.event_df[i].end_time.values[0]
dur = etime - stime
if dur < duration_thresh:
print 'Stim %d (%s) is too short: %0.3fs' % (stim_id, stim_type, dur)
continue
si = int(stime * rp.sample_rate)
ei = int(etime * rp.sample_rate)
u = stim_env[si:ei]
env_freq, env_ps, env_ps_var, env_phase = power_spectrum_jn(u, rp.sample_rate, window_length=1.0,
increment=0.200)
env_data['stim_id'].append(stim_id)
env_data['stim_type'].append(stim_type)
env_data['duration'].append(etime - stime)
env_data['xindex'].append(len(env_psds))
env_psds.append(env_ps)
env_psds = np.array(env_psds)
fi = (env_freq >= 2.) & (env_freq <= 30.)
env_psds = env_psds[:, fi]
env_freq = env_freq[fi]
env_df = pd.DataFrame(env_data)
return env_freq,env_psds,env_df
def testFFT(self):
sr = 1000.
freqs = [35.]
dur = 0.500
nt = int(dur * sr)
t = np.arange(nt) / sr
# create a psth that has the specific frequencies
psth = np.zeros([nt])
for f in freqs:
psth += np.sin(2 * np.pi * f * t)
max_spike_rate = 0.1
psth /= psth.max()
psth += 1.
psth /= 2.0
psth *= max_spike_rate
# simulate a spike train with a variety of frequencies in it
trials = simulate_poisson(psth, dur, num_trials=10)
bin_size = 0.001
binned_trials = spike_trains_to_matrix(trials, bin_size, 0.0, dur)
mean_psth = binned_trials.mean(axis=0)
# compute the power spectrum of each spike train
psds = list()
pfreq = None
win_len = 0.090
inc = 0.010
for st in binned_trials:
pfreq, psd, ps_var, phase = power_spectrum_jn(
st, 1.0 / bin_size, win_len, inc)
nz = psd > 0
psd[nz] = 20 * np.log10(psd[nz]) + 100
psd[psd < 0] = 0
psds.append(psd)
psds = np.array(psds)
mean_psd = psds.mean(axis=0)
pfreq, mean_psd2, ps_var, phase = power_spectrum_jn(
mean_psth, 1.0 / bin_size, win_len, inc)
nz = mean_psd2 > 0
mean_psd2[nz] = 20 * np.log10(mean_psd2[nz]) + 100
mean_psd2[mean_psd2 < 0] = 0
plt.figure()
ax = plt.subplot(2, 1, 1)
plot_raster(trials,
ax=ax,
duration=dur,
bin_size=0.001,
time_offset=0.0,
ylabel='Trial #',
bgcolor=None,
spike_color='k')
ax = plt.subplot(2, 1, 2)
plt.plot(pfreq, mean_psd, 'k-', linewidth=3.0)
for psd in psds:
plt.plot(pfreq, psd, '-', linewidth=2.0, alpha=0.75)
plt.plot(pfreq, mean_psd2, 'k--', linewidth=3.0, alpha=0.60)
plt.axis('tight')
plt.xlabel('Frequency (Hz)')
plt.ylabel('Power (dB)')
plt.xlim(0, 100.)
plt.show()
def testFFT(self):
sr = 1000.
freqs = [35.]
dur = 0.500
nt = int(dur*sr)
t = np.arange(nt) / sr
# create a psth that has the specific frequencies
psth = np.zeros([nt])
for f in freqs:
psth += np.sin(2*np.pi*f*t)
max_spike_rate = 0.1
psth /= psth.max()
psth += 1.
psth /= 2.0
psth *= max_spike_rate
# simulate a spike train with a variety of frequencies in it
trials = simulate_poisson(psth, dur, num_trials=10)
bin_size = 0.001
binned_trials = spike_trains_to_matrix(trials, bin_size, 0.0, dur)
mean_psth = binned_trials.mean(axis=0)
# compute the power spectrum of each spike train
psds = list()
pfreq = None
win_len = 0.090
inc = 0.010
for st in binned_trials:
pfreq,psd,ps_var,phase = power_spectrum_jn(st, 1.0 / bin_size, win_len, inc)
nz = psd > 0
psd[nz] = 20*np.log10(psd[nz]) + 100
psd[psd < 0] = 0
psds.append(psd)
psds = np.array(psds)
mean_psd = psds.mean(axis=0)
pfreq,mean_psd2,ps_var,phase = power_spectrum_jn(mean_psth, 1.0/bin_size, win_len, inc)
nz = mean_psd2 > 0
mean_psd2[nz] = 20*np.log10(mean_psd2[nz]) + 100
mean_psd2[mean_psd2 < 0] = 0
plt.figure()
ax = plt.subplot(2, 1, 1)
plot_raster(trials, ax=ax, duration=dur, bin_size=0.001, time_offset=0.0, ylabel='Trial #', bgcolor=None, spike_color='k')
ax = plt.subplot(2, 1, 2)
plt.plot(pfreq, mean_psd, 'k-', linewidth=3.0)
for psd in psds:
plt.plot(pfreq, psd, '-', linewidth=2.0, alpha=0.75)
plt.plot(pfreq, mean_psd2, 'k--', linewidth=3.0, alpha=0.60)
plt.axis('tight')
plt.xlabel('Frequency (Hz)')
plt.ylabel('Power (dB)')
plt.xlim(0, 100.)
plt.show()
def compute_spectra_and_coherence_single_electrode(lfp1, lfp2, sample_rate, e1, e2,
window_length=0.060, increment=None, log=True,
window_fraction=0.60, noise_floor_db=25,
lags=np.arange(-20, 21, 1), psd_stats=None):
"""
:param lfp1: An array of shape (ntrials, nt)
:param lfp2: An array of shape (ntrials, nt)
:return:
"""
# compute the mean (locked) spectra
lfp1_mean = lfp1.mean(axis=0)
lfp2_mean = lfp2.mean(axis=0)
if increment is None:
increment = 2.0 / sample_rate
pfreq,psd1,ps_var,phase = power_spectrum_jn(lfp1_mean, sample_rate, window_length, increment)
pfreq,psd2,ps_var,phase = power_spectrum_jn(lfp2_mean, sample_rate, window_length, increment)
if log:
log_transform(psd1)
log_transform(psd2)
c12 = coherency(lfp1_mean, lfp2_mean, lags, window_fraction=window_fraction, noise_floor_db=noise_floor_db)
# compute the nonlocked spectra coherence
c12_pertrial = list()
ntrials,nt = lfp1.shape
psd1_ms_all = list()
psd2_ms_all = list()
for k in range(ntrials):
i = np.ones([ntrials], dtype='bool')
i[k] = False
lfp1_jn_mean = lfp1[i, :].mean(axis=0)
lfp2_jn_mean = lfp2[i, :].mean(axis=0)
lfp1_ms = lfp1[k, :] - lfp1_jn_mean
lfp2_ms = lfp2[k, :] - lfp2_jn_mean
pfreq,psd1_ms,ps_var_ms,phase_ms = power_spectrum_jn(lfp1_ms, sample_rate, window_length, increment)
pfreq,psd2_ms,ps_var_ms,phase_ms = power_spectrum_jn(lfp2_ms, sample_rate, window_length, increment)
if log:
log_transform(psd1_ms)
log_transform(psd2_ms)
psd1_ms_all.append(psd1_ms)
psd2_ms_all.append(psd2_ms)
c12_ms = coherency(lfp1_ms, lfp2_ms, lags, window_fraction=window_fraction, noise_floor_db=noise_floor_db)
c12_pertrial.append(c12_ms)
psd1_ms_all = np.array(psd1_ms_all)
psd2_ms_all = np.array(psd2_ms_all)
psd1_ms = psd1_ms_all.mean(axis=0)
psd2_ms = psd2_ms_all.mean(axis=0)
if psd_stats is not None:
psd_mean1,psd_std1 = psd_stats[e1]
psd_mean2,psd_std2 = psd_stats[e2]
psd1 -= psd_mean1
psd1 /= psd_std1
psd2 -= psd_mean2
psd2 /= psd_std2
psd1_ms -= psd_mean1
psd1_ms /= psd_std1
psd2_ms -= psd_mean2
psd2_ms /= psd_std2
c12_pertrial = np.array(c12_pertrial)
c12_nonlocked = c12_pertrial.mean(axis=0)
# compute the coherence per trial then take the average
c12_totals = list()
for k in range(ntrials):
c12 = coherency(lfp1[k, :], lfp2[k, :], lags, window_fraction=window_fraction, noise_floor_db=noise_floor_db)
c12_totals.append(c12)
c12_totals = np.array(c12_totals)
c12_total = c12_totals.mean(axis=0)
return pfreq, psd1, psd2, psd1_ms, psd2_ms, c12, c12_nonlocked, c12_total
def compute_spectra_and_coherence_multi_electrode_single_trial(lfps, sample_rate, electrode_indices, electrode_order,
window_length=0.060, increment=None, log=True,
window_fraction=0.60, noise_floor_db=25,
lags=np.arange(-20, 21, 1),
psd_stats=None):
"""
:param lfps: an array of shape (ntrials, nelectrodes, nt)
:return:
"""
if increment is None:
increment = 2.0 / sample_rate
nelectrodes,nt = lfps.shape
freqs = get_freqs(sample_rate, window_length, increment)
lags_ms = get_lags_ms(sample_rate, lags)
spectra = np.zeros([nelectrodes, len(freqs)])
cross_mat = np.zeros([nelectrodes, nelectrodes, len(lags_ms)])
for k in range(nelectrodes):
_e1 = electrode_indices[k]
i1 = electrode_order.index(_e1)
lfp1 = lfps[k, :]
freqs,psd1,ps_var,phase = power_spectrum_jn(lfp1, sample_rate, window_length, increment)
if log:
log_transform(psd1)
if psd_stats is not None:
psd_mean,psd_std = psd_stats[_e1]
"""
plt.figure()
plt.subplot(2, 2, 1)
plt.plot(freqs, psd1, 'k-')
plt.title('PSD (%d)' % _e1)
plt.axis('tight')
plt.subplot(2, 2, 3)
plt.plot(freqs, psd_mean, 'g-')
plt.title('Mean')
plt.axis('tight')
plt.subplot(2, 2, 4)
plt.plot(freqs, psd_std, 'c-')
plt.title('STD')
plt.axis('tight')
plt.subplot(2, 2, 2)
psd1_z = deepcopy(psd1)
psd1_z -= psd_mean
psd1_z /= psd_std
plt.plot(freqs, psd1_z, 'r-')
plt.title('Zscored')
plt.axis('tight')
"""
psd1 -= psd_mean
psd1 /= psd_std
spectra[i1, :] = psd1
for j in range(k):
_e2 = electrode_indices[j]
i2 = electrode_order.index(_e2)
lfp2 = lfps[j, :]
cf = coherency(lfp1, lfp2, lags, window_fraction=window_fraction, noise_floor_db=noise_floor_db)
"""
freqs,c12,c_var_amp,c_phase,c_phase_var,coherency,coherency_t = coherence_jn(lfp1, lfp2, sample_rate,
window_length, increment,
return_coherency=True)
"""
cross_mat[i1, i2] = cf
cross_mat[i2, i1] = cf[::-1]
return spectra, cross_mat