low = 0.5
high = 2
cells, _ = parse_cellid(ops)
rec = generate_state_corrected_psth(batch=batch,
modelname=xmodel,
cellids=cells,
siteid=site,
cache_path=path,
recache=False)
ff_residuals = rec['resp']._data - rec['psth_sp']._data
nan_idx = np.isnan(ff_residuals[0, :])
ff_residuals[:, nan_idx] = 0
ff_residuals = bandpass_filter_resp(ff_residuals,
low,
high,
fs=fs,
boxcar=True)
rec['ff_residuals'] = rec['resp']._modified_copy(ff_residuals)
rec = rec.apply_mask()
pupil = rec['pupil']._data
raw_residual = rec['resp']._data - rec['psth_sp']._data
cor_residual = rec['resp']._data - rec['psth']._data
# first, do full PCA on residuals
pca = PCA()
pca.fit(raw_residual.T)
pca_transform = raw_residual.T.dot(pca.components_.T).T
# do first order regression
'Regress first order pupil by subtracting model pred using pupil pred alone'
)
# set the LV to 0 by setting encoding weights to 0 and compute prediction
ms = ctx['modelspec']
ms.phi[1]['e'] = np.zeros(ms.phi[1]['e'].shape)
pred = ms.evaluate(rec)
mod_data = rec['resp']._data - pred['pred']._data + rec[
'psth_sp']._data
rec['resp'] = rec['resp']._modified_copy(mod_data)
else:
raise ValueError("No regression method specified!")
if filt:
log.info("Band-pass filter spike counts between {0} and {1} Hz".format(
low_c, high_c))
rec = preproc.bandpass_filter_resp(rec, low_c, high_c)
# USE XFORMS PUPIL MASK
rec_bp = rec.copy()
rec_bp['mask'] = rec_bp['p_mask']
rec_sp = rec.copy()
rec_sp['mask'] = rec_sp['p_mask']._modified_copy(~rec_sp['p_mask']._data)
rec_bp = rec_bp.apply_mask(reset_epochs=True)
rec_sp = rec_sp.apply_mask(reset_epochs=True)
# Determine if need to mask pupil balanced epochs only, or if the
# xforms model already did this
if not balanced:
if batch == 289:
balanced_eps = preproc.get_pupil_balanced_epochs(
pca = PCA()
pca.fit(raw_residual.T)
f, ax = plt.subplots(1, 3, figsize=(12, 4))
# scree plot
ax[0].plot(pca.explained_variance_ratio_, 'o-', label='raw')
ax[0].set_ylabel('% Variance explained')
ax[0].set_xlabel('PC')
ax[0].legend(frameon=False)
# correlation with pupil
transform = raw_residual.T.dot(pca.components_.T)
shuf_transform = shuff_residual.T.dot(pca.components_.T)
corr = [abs(np.corrcoef(transform[:, i], pupil)[0, 1]) for i in range(residual.shape[0])]
corr_noise_floor = [abs(np.corrcoef(shuf_transform[:, i], pupil)[0, 1]) for i in range(residual.shape[0])]
ax[1].plot(corr, 'o-', label='raw')
ax[1].plot(corr_noise_floor, 'o-', label='shuffled')
ax[1].set_ylabel('Corr. Coef w/ pupil')
ax[1].set_xlabel('PC')
ax[1].legend(frameon=False)
# correlation of power with pupil
transform_filter = bandpass_filter_resp(transform.T, low, high, fs=100)
# TODO convert to power over sliding window...
f.tight_layout()
plt.show()
# compute PCs over all the data
pca = PCA()
pca.fit(resp_matrix.T)
# compute variance of rank1 matrix along each PC
var = np.zeros(resp_matrix.shape[0])
fo_var = np.zeros(pred_matrix.shape[0])
for pc in range(0, resp_matrix.shape[0]):
var[pc] = np.var(r1_resp.T.dot(pca.components_[pc])) / np.sum(pca.explained_variance_)
fo_var[pc] = np.var(pred_matrix.T.dot(pca.components_[pc])) / np.sum(pca.explained_variance_)
# project onto first evec and filter
#rm = cpreproc.bandpass_filter_resp(resp_matrix, low_c=low, high_c=high, fs=fs)
proj = resp_matrix.T.dot(evecs[:, 0])[np.newaxis]
proj_filt = proj.copy()
proj_filt = cpreproc.bandpass_filter_resp(proj, low_c=low, high_c=high, fs=fs)
# caculate power in the filtered residual
t, ffsw = sliding_window(proj_filt, fs=fs, window_size=4, step_size=1)
power = np.sum(ffsw**2, axis=-1) / ffsw.shape[-1]
t, pds = sliding_window(pupil[np.newaxis, :], fs=fs, window_size=4, step_size=1)
pds = pds.mean(axis=-1)
# plot results
ncells = resp_matrix.shape[0]
f = plt.figure(figsize=(15, 6))
espec = plt.subplot2grid((2, 5), (0, 0))
pcvar = plt.subplot2grid((2, 5), (0, 3), colspan=2)
pax = plt.subplot2grid((2, 5), (1, 0), colspan=3)
betaax = plt.subplot2grid((2, 5), (0, 1))