eigvals = np.zeros((st.nclusters, stimdim, fl))
eigvecs = np.zeros((st.nclusters, stimdim, fl, fl))
cross_corrs = np.zeros(st.nclusters)
for i, cl in enumerate(st.clusters):
sta = data['stas'][i]
spikes = st.binnedspiketimes(i)
import time
start = time.time()
# Calculate the contrast for each cell's receptive field
stimulus[-1, :] = st.contrast_signal_cell(i)
sp_tr, sp_te, stim_tr, stim_te = train_test_split(spikes, stimulus,
test_size=val_split_size,
split_pos=val_split_pos)
res = gqm.minimize_loglikelihood(np.zeros((stimdim, fl)),
np.zeros((stimdim, fl, fl)), 0,
stim_tr,
st.frame_duration,
sp_tr,
minimize_disp=True,
method='BFGS')
elapsed = time.time()-start
print(f'Time elapsed: {elapsed/60:6.1f} mins for cell {i}')
t = np.linspace(0, st.filter_length * st.frame_duration * 1000,
st.filter_length)
plotlabels = ['Contrast']
# Temporary hack until FFF gets its own class.
fff.readpars()
if fff.param_file['Nblinks'] == 1:
t_fff = np.linspace(0, st.filter_length * st.frame_duration * 1000,
st.filter_length * 2)
else:
t_fff = t
for i, cluster in enumerate(st.clusters):
stimulus = st.contrast_signal_cell(i).squeeze()
res = glm.minimize_loglhd(np.zeros(st.filter_length),
0,
stimulus,
st.frame_duration,
all_spikes[i, :],
usegrad=True,
method='BFGS')
if not res['success']:
print(i, 'did not complete successfully.')
kall[i, :] = res['x'][:-1]
muall[i] = res['x'][-1]
frs[i, :] = glm.glm_fr(kall[i, :], muall[i], st.frame_duration)(stimulus)
cross_corrs[i] = np.corrcoef(all_spikes[i, :], frs[i, :])[0, 1]
def omb_contrastmotion2dnonlin(exp,
stim,
nbins_nlt=9,
cmap='Greys',
plot3d=False):
"""
Calculate and plot the 2D nonlinearities for the OMB stimulus. The
magnitude of the stimulus projection on quadratic motion filters
from GQM is used for the motion.
Parameters:
------
nbins_nlt:
Number of bins to be used for dividing the generator signals
into ranges with equal number of samples.
plot3d:
Whether to additionally create a 3D version of the nonlinearity.
"""
st = OMB(exp, stim)
# Motion and contrast
data_cm = np.load(
os.path.join(st.exp_dir, 'data_analysis', st.stimname,
'GQM_motioncontrast', f'{stim}_GQM_motioncontrast.npz'))
qall = data_cm['Qall']
kall = data_cm['kall']
muall = data_cm['muall']
cross_corrs = data_cm['cross_corrs']
allspikes = st.allspikes()
stim_mot = st.bgsteps.copy()
# Bin dimension should be one greater than nonlinearity for pcolormesh
# compatibility. Otherwise the last row and column of nonlinearity is not
# plotted.
all_bins_c = np.zeros((st.nclusters, nbins_nlt + 1))
all_bins_r = np.zeros((st.nclusters, nbins_nlt + 1))
nonlinearities = np.zeros((st.nclusters, nbins_nlt, nbins_nlt))
label = '2D-nonlin_magQ_motion_kcontrast'
savedir = os.path.join(st.stim_dir, label)
os.makedirs(savedir, exist_ok=True)
for i in range(st.nclusters):
stim_con = st.contrast_signal_cell(i).squeeze()
# Project the motion stimulus onto the quadratic filter
generator_x = gqm.conv2d(qall[i, 0, :], stim_mot[0, :])
generator_y = gqm.conv2d(qall[i, 1, :], stim_mot[1, :])
# Calculate the magnitude of the vector formed by motion generators
generators = np.vstack([generator_x, generator_y])
r = np.sqrt(np.sum(generators**2, axis=0))
# Project the contrast stimulus onto the linear filter
generator_c = np.convolve(stim_con, kall[i, 2, :],
'full')[:-st.filter_length + 1]
spikes = allspikes[i, :]
nonlinearity, bins_c, bins_r = nlt.calc_nonlin_2d(spikes,
generator_c,
r,
nr_bins=nbins_nlt)
nonlinearity /= st.frame_duration
all_bins_c[i, :] = bins_c
all_bins_r[i, :] = bins_r
nonlinearities[i, ...] = nonlinearity
X, Y = np.meshgrid(bins_c, bins_r, indexing='ij')
fig = plt.figure()
gs = gsp.GridSpec(5, 5)
axmain = plt.subplot(gs[1:, :-1])
axx = plt.subplot(gs[0, :-1], sharex=axmain)
axy = plt.subplot(gs[1:, -1], sharey=axmain)
# Normally subplots turns off shared axis tick labels but
# Gridspec does not do this
plt.setp(axx.get_xticklabels(), visible=False)
plt.setp(axy.get_yticklabels(), visible=False)
im = axmain.pcolormesh(X, Y, nonlinearity, cmap=cmap)
plf.integerticks(axmain)
cb = plt.colorbar(im)
cb.outline.set_linewidth(0)
cb.ax.set_xlabel('spikes/s')
cb.ax.xaxis.set_label_position('top')
plf.integerticks(cb.ax, 4, which='y')
plf.integerticks(axx, 1, which='y')
plf.integerticks(axy, 1, which='x')
barkwargs = dict(alpha=.3, facecolor='k', linewidth=.5, edgecolor='w')
axx.bar(nlt.bin_midpoints(bins_c),
nonlinearity.mean(axis=1),
width=np.ediff1d(bins_c),
**barkwargs)
axy.barh(nlt.bin_midpoints(bins_r),
nonlinearity.mean(axis=0),
height=np.ediff1d(bins_r),
**barkwargs)
plf.spineless(axx, 'b')
plf.spineless(axy, 'l')
axmain.set_xlabel('Projection onto linear contrast filter')
axmain.set_ylabel(
'Magnitude of projection onto quadratic motion filters')
fig.suptitle(
f'{st.exp_foldername}\n{st.stimname}\n{st.clids[i]} '
f'2D nonlinearity nsp: {st.allspikes()[i, :].sum():<5.0f}')
plt.subplots_adjust(top=.85)
fig.savefig(os.path.join(savedir, st.clids[i]), bbox_inches='tight')
plt.show()
if plot3d:
if i == 0:
from mpl_toolkits import mplot3d
from matplotlib.ticker import MaxNLocator
#%%
fig = plt.figure()
ax = plt.axes(projection='3d')
ax.plot_surface(X,
Y,
nonlinearity,
cmap='YlGn',
edgecolors='k',
linewidths=0.2)
ax.set_xlabel('Projection onto linear contrast filter')
ax.set_ylabel(
'Magnitude of projection onto quadratic motion filters')
ax.set_zlabel(r'Firing rate [sp/s]')
ax.view_init(elev=30, azim=-135)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.zaxis.set_major_locator(MaxNLocator(integer=True))
keystosave = ['nonlinearities', 'all_bins_c', 'all_bins_r', 'nbins_nlt']
datadict = {}
for key in keystosave:
datadict.update({key: locals()[key]})
npzfpath = os.path.join(savedir, f'{st.stimnr}_{label}.npz')
np.savez(npzfpath, **datadict)
def omb_contrastmotion2dnonlin_Qcomps(exp, stim, nbins_nlt=9, cmap='Greys'):
"""
Calculate and plot the 2D nonlinearities for the OMB stimulus. Multiple
components of the matrix Q for the motion.
Parameters:
------
nbins_nlt:
Number of bins to be used for dividing the generator signals
into ranges with equal number of samples.
"""
st = OMB(exp, stim)
# Motion and contrast
data_cm = np.load(os.path.join(st.exp_dir, 'data_analysis',
st.stimname, 'GQM_motioncontrast_val',
f'{stim}_GQM_motioncontrast_val.npz'))
qall = data_cm['Qall']
kall = data_cm['kall']
muall = data_cm['muall']
eigvecs = data_cm['eigvecs']
eigvals = data_cm['eigvals']
eiginds = [-1, 0] # activating, suppressing #HINT
cross_corrs = data_cm['cross_corrs']
allspikes = st.allspikes()
stim_mot = st.bgsteps.copy()
# Bin dimension should be one greater than nonlinearity for pcolormesh
# compatibility. Otherwise the last row and column of nonlinearity is not
# plotted.
all_bins_c = np.zeros((st.nclusters, nbins_nlt+1))
all_bins_r = np.zeros((st.nclusters, nbins_nlt+1))
nonlinearities = np.zeros((st.nclusters, nbins_nlt, nbins_nlt))
label = '2D-nonlin_Qallcomps_motion_kcontrast'
row_labels = ['Activating', 'Suppresive']
column_labels = ['X', 'Y', r'$\sqrt{X^2 + Y^2}$']
savedir = os.path.join(st.stim_dir, label)
os.makedirs(savedir, exist_ok=True)
for i in range(st.nclusters):
stim_con = st.contrast_signal_cell(i).squeeze()
n = 3 # x, y, xy
m = 2 # activating, suppressing
fig = plt.figure(figsize=(n*5, m*5), constrained_layout=True)
gs = fig.add_gridspec(m, n)
axes = []
for _, eachgs in enumerate(gs):
subgs = eachgs.subgridspec(2, 3, width_ratios=[4, 1, .2], height_ratios=[1, 4])
mainax = fig.add_subplot(subgs[1, 0])
axx = fig.add_subplot(subgs[0, 0], sharex=mainax)
axy = fig.add_subplot(subgs[1, 1], sharey=mainax)
cbax = fig.add_subplot(subgs[1, 2])
axes.append([axx, mainax, axy, cbax])
for k, eigind in enumerate(eiginds):
generator_x = np.convolve(eigvecs[i, 0, :, eigind],
stim_mot[0, :], 'full')[:-st.filter_length+1]
generator_y = np.convolve(eigvecs[i, 1, :, eigind],
stim_mot[1, :], 'full')[:-st.filter_length+1]
# Calculate the magnitude of the vector formed by motion generators
generators = np.vstack([generator_x, generator_y])
generator_xy = np.sqrt(np.sum(generators**2, axis=0))
# Project the contrast stimulus onto the linear filter
generator_c = np.convolve(stim_con,
kall[i, 2, :],
'full')[:-st.filter_length+1]
spikes = allspikes[i, :]
generators_motion = [generator_x, generator_y, generator_xy]
for l, direction in enumerate(column_labels):
nonlinearity, bins_c, bins_r = nlt.calc_nonlin_2d(spikes,
generator_c,
generators_motion[l],
nr_bins=nbins_nlt)
nonlinearity /= st.frame_duration
all_bins_c[i, :] = bins_c
all_bins_r[i, :] = bins_r
nonlinearities[i, ...] = nonlinearity
X, Y = np.meshgrid(bins_c, bins_r, indexing='ij')
subaxes = axes[k*n+l]
axmain = subaxes[1]
axx = subaxes[0]
axy = subaxes[2]
cbax = subaxes[3]
# Normally subplots turns off shared axis tick labels but
# Gridspec does not do this
plt.setp(axx.get_xticklabels(), visible=False)
plt.setp(axy.get_yticklabels(), visible=False)
im = axmain.pcolormesh(X, Y, nonlinearity, cmap=cmap)
plf.integerticks(axmain, 6, which='xy')
cb = plt.colorbar(im, cax=cbax)
cb.outline.set_linewidth(0)
cb.ax.set_xlabel('spikes/s')
cb.ax.xaxis.set_label_position('top')
plf.integerticks(cb.ax, 4, which='y')
plf.integerticks(axx, 1, which='y')
plf.integerticks(axy, 1, which='x')
barkwargs = dict(alpha=.3, facecolor='k',
linewidth=.5, edgecolor='w')
axx.bar(nlt.bin_midpoints(bins_c), nonlinearity.mean(axis=1),
width=np.ediff1d(bins_c), **barkwargs)
axy.barh(nlt.bin_midpoints(bins_r), nonlinearity.mean(axis=0),
height=np.ediff1d(bins_r), **barkwargs)
plf.spineless(axx, 'b')
plf.spineless(axy, 'l')
if k == 0 and l == 0:
axmain.set_xlabel('Projection onto linear contrast filter')
axmain.set_ylabel(f'Projection onto Q component')
if k == 0:
axx.set_title(direction)
if l == 0:
axmain.text(-.3, .5, row_labels[k],
va='center',
rotation=90,
transform=axmain.transAxes)
fig.suptitle(f'{st.exp_foldername}\n{st.stimname}\n{st.clids[i]} '
f'2D nonlinearity nsp: {st.allspikes()[i, :].sum():<5.0f}')
plt.subplots_adjust(top=.85)
fig.savefig(os.path.join(savedir, st.clids[i]), bbox_inches='tight')
plt.show()
keystosave = ['nonlinearities', 'all_bins_c', 'all_bins_r', 'nbins_nlt']
datadict = {}
for key in keystosave:
datadict.update({key: locals()[key]})
npzfpath = os.path.join(savedir, f'{st.stimnr}_{label}.npz')
np.savez(npzfpath, **datadict)
