params = gpcsd_model.extract_model_params()
with open('%s/results/gpcsd_model_%s.pkl' % (root_path, probe_name),
'wb') as f:
pickle.dump(params, f)
# %% Predict during baseline period
gpcsd_model.update_lfp(lfp_baseline, t)
print(gpcsd_model)
gpcsd_model.predict(x, t)
tcsd = predictcsd_trad_1d(lfp_baseline)
# %% Single trial image plots
trial = 0
plt.figure(figsize=(14, 5))
plt.subplot(151)
plt.imshow(normalize(gpcsd_model.csd_pred_list[0][:, :, trial]),
aspect='auto',
cmap='bwr',
vmin=-1,
vmax=1)
plt.title('Slow CSD')
plt.subplot(152)
plt.imshow(normalize(gpcsd_model.csd_pred_list[1][:, :, trial]),
aspect='auto',
cmap='bwr',
vmin=-1,
vmax=1)
plt.title('Fast CSD')
plt.subplot(153)
plt.imshow(normalize(gpcsd_model.csd_pred[:, :, trial]),
aspect='auto',
csd3_interior_electrodes = np.zeros((nx - 2, nt, ntrials))
for trial in range(ntrials):
csdinterp = scipy.interpolate.RectBivariateSpline(z, t, csd2[:, :, trial])
csd2_interior_electrodes[:, :, trial] = csdinterp(xshort, t)
csdinterp = scipy.interpolate.RectBivariateSpline(z, t, csd3[:, :, trial])
csd3_interior_electrodes[:, :, trial] = csdinterp(xshort, t)
# %% Pass through forward model, add white noise
lfp2 = np.zeros((nx, nt, ntrials))
lfp3 = np.zeros((nx, nt, ntrials))
for trial in range(ntrials):
lfp2[:, :, trial] = fwd_model_1d(csd2[:, :, trial], z, x, 100)
lfp3[:, :, trial] = fwd_model_1d(csd3[:, :, trial], z, x, 100)
lfp2 = lfp2 + np.random.normal(0, np.sqrt(0.0001), size=(nx, nt, ntrials))
lfp3 = lfp3 + np.random.normal(0, np.sqrt(0.0001), size=(nx, nt, ntrials))
lfp2 = normalize(lfp2)
lfp3 = normalize(lfp3)
# %%
plt.figure(figsize=(8, 5))
plt.subplot(121)
plt.imshow(normalize(csd2[:, :, 0]),
vmin=-1,
vmax=1,
cmap='bwr',
aspect='auto')
plt.xlabel('Time')
plt.ylabel('Depth')
plt.title('Two-component CSD')
plt.colorbar()
plt.subplot(122)
gpcsd_model.fit(n_restarts=10, verbose=True)
# %% predict
print(gpcsd_model)
gpcsd_model.predict(z_grid, t)
csd_pred_rect = gpcsd_model.csd_pred.reshape((nz1, nz2, nt, -1))
# %% trad CSD on test data
tcsd_pred = predictcsd_trad_2d(lfp_sparse_rect)[:, 1:-1, :, :]
# %%
tinds = [0, 1, 2, 3]
plt.figure(figsize=(16, 16))
for ti in tinds:
plt.subplot(4, 4, ti + 1)
plt.imshow(normalize(tcsd_pred[:, :, ti, 0]).T, aspect='auto', cmap='bwr')
for ti in tinds:
plt.subplot(4, 4, ti + 1 + 4)
plt.imshow(normalize(csd_pred_rect[:, :, ti, 0]).T,
aspect='auto',
cmap='bwr')
for ti in tinds:
plt.subplot(4, 4, ti + 1 + 8)
plt.imshow(normalize(lfp_sparse_rect[:, :, ti, 0]).T,
aspect='auto',
cmap='bwr')
# %% Compute RMSE and R2
rmse = np.sqrt(np.mean(np.square(csd_pred_rect - csd_dense_rect)))
evoked_probe['kcsd'] = kcsd_evoked_model.values()
with open('%s/results/gpcsd_evoked_%s.pkl' % (root_path, probe_name),
'wb') as f:
pickle.dump(evoked_probe, f)
evoked[probe_name] = evoked_probe
with open('%s/results/gpcsd_cov_%s.pkl' % (root_path, probe_name),
'wb') as f:
pickle.dump(cov_probe, f)
cov[probe_name] = cov_probe
# Supplemental figure: comparing kCSD on evoked and GPCSD evoked
plt.figure(figsize=(14, 6))
plt.subplot(131)
plt.imshow(normalize(evoked[probe_name]['lfp']),
vmin=-1,
vmax=1,
aspect='auto',
cmap='bwr')
plt.title('%s evoked LFP' % probe_name)
plt.subplot(132)
plt.imshow(normalize(evoked[probe_name]['gpcsd']),
vmin=-1,
vmax=1,
aspect='auto',
cmap='bwr')
plt.title('GPCSD evoked')
plt.subplot(133)
plt.imshow(normalize(evoked[probe_name]['kcsd']),
cmap='bwr',
gpcsd_gen.temporal_cov_list[1].params['sigma2']['value'] = sig2tM_true
# %% Generate CSD and sample at interior electrode positions for comparing to tCSD
csd = gpcsd_gen.sample_prior(2 * ntrials)
csd_interior_electrodes = np.zeros((nx - 2, nt, 2 * ntrials))
for trial in range(2 * ntrials):
csdinterp = scipy.interpolate.RectBivariateSpline(z, t, csd[:, :, trial])
csd_interior_electrodes[:, :, trial] = csdinterp(xshort, t)
# %% Pass through forward model, add white noise
lfp = np.zeros((nx, nt, 2 * ntrials))
for trial in range(2 * ntrials):
lfp[:, :, trial] = fwd_model_1d(csd[:, :, trial], z, x, R_true)
lfp = lfp + np.random.normal(
0, np.sqrt(sig2n_true), size=(nx, nt, 2 * ntrials))
lfp = normalize(lfp)
# %% Visualize one trial
plt.figure()
plt.subplot(121)
plt.imshow(csd[:, :, 0], vmin=-1, vmax=1, cmap='bwr', aspect='auto')
plt.title('CSD')
plt.xlabel('Time')
plt.ylabel('depth')
plt.colorbar()
plt.subplot(122)
plt.imshow(lfp[:, :, 0], cmap='bwr', aspect='auto')
plt.title('LFP')
plt.xlabel('Time')
plt.colorbar()
plt.tight_layout()
nx = 24 # Number of observed spatial locations for LFP
snr = 30 # noise characteristics
t = np.linspace(0, 50, nt)[:, None]
x = np.linspace(a, b, nx)[:, None]
nz = int(np.rint(b - a) / deltaz) + 1
z = np.linspace(a, b, nz)[:, None]
# %% Create true CSD
csd_true = csd_true_f(z, t)
print('generated CSD of shape')
print(csd_true.shape)
# %% Create observed LFP
lfp = {}
lfp['noiseless'] = normalize(
forward_models.fwd_model_1d(csd_true, z, x, R_true))
print('generated LFP of shape')
print(lfp['noiseless'].shape)
# Add white noise to LFP
sig2n_true = np.square(np.std(lfp['noiseless']) / snr)
lfp['white_noise'] = lfp['noiseless'] + np.random.normal(
0, np.sqrt(sig2n_true), size=(nx, nt))
plt.figure(figsize=(12, 6))
plt.subplot(131)
plt.imshow(csd_true, aspect='auto', cmap='bwr')
plt.title('True CSD')
plt.ylabel('Depth (microns)')
plt.xlabel('Time (ms)')
plt.subplot(132)