def psth_per_file(rec):
raise NotImplementedError
resp = rec['resp'].rasterize()
file_epochs = ep.epoch_names_matching(resp.epochs, "^FILE_")
epoch_regex = "^STIM_"
stim_epochs = ep.epoch_names_matching(resp.epochs, epoch_regex)
r = []
max_rep_id = np.zeros(len(file_epochs))
for f in file_epochs:
r.append(resp.as_matrix(stim_epochs, overlapping_epoch=f) * resp.fs)
repcount = np.sum(np.isfinite(r[:, :, 0, 0]), axis=1)
max_rep_id, = np.where(repcount == np.max(repcount))
t = np.arange(r.shape[-1]) / resp.fs
plt.figure()
ax = plt.subplot(3, 1, 1)
nplt.plot_spectrogram(s[max_rep_id[-1], 0, :, :],
fs=stim.fs,
ax=ax,
title="cell {} - stim".format(cellid))
ax = plt.subplot(3, 1, 2)
nplt.raster(t, r[max_rep_id[-1], :, 0, :], ax=ax, title='raster')
ax = plt.subplot(3, 1, 3)
nplt.psth_from_raster(t,
r[max_rep_id[-1], :, 0, :],
ax=ax,
title='raster',
ylabel='spk/s')
plt.tight_layout()
def test_plots():
recording.get_demo_recordings(name=recording_file)
rec = recording.load_recording(uri)
resp = rec['resp'].rasterize()
stim = rec['stim'].rasterize()
epoch_regex = "^STIM_"
stim_epochs = ep.epoch_names_matching(resp.epochs, epoch_regex)
r = resp.as_matrix(stim_epochs) * resp.fs
s = stim.as_matrix(stim_epochs)
repcount = np.sum(np.isfinite(r[:, :, 0, 0]), axis=1)
max_rep_id, = np.where(repcount == np.max(repcount))
t = np.arange(r.shape[-1]) / resp.fs
plt.figure()
ax = plt.subplot(3, 1, 1)
nplt.plot_spectrogram(s[max_rep_id[-1], 0, :, :],
fs=stim.fs,
ax=ax,
title="cell {} - stim".format(cellid))
ax = plt.subplot(3, 1, 2)
nplt.raster(t, r[max_rep_id[-1], :, 0, :], ax=ax, title='raster')
ax = plt.subplot(3, 1, 3)
nplt.psth_from_raster(t,
r[max_rep_id[-1], :, 0, :],
ax=ax,
title='raster',
ylabel='spk/s')
plt.tight_layout()
def model_pred_comp(cellid, batch, modelnames, occurrence=None,
pre_dur=None, dur=None):
"""
return ccd12, ccd13, ccd23
"""
modelcount = len(modelnames)
epoch = 'REFERENCE'
c = 0
plot_colors = ['lightgray', 'g', 'lightgray', 'r', 'lightgray', 'b']
legend = ['act','LN','act','GC','act','STP']
times = []
values = []
r_test = []
for i, m in enumerate(modelnames):
xf, ctx = load_model_baphy_xform(cellid, batch, m)
val = ctx['val'][0]
if i == 0:
d = val['resp'].get_epoch_bounds('PreStimSilence')
if len(d):
PreStimSilence = np.mean(np.diff(d))
else:
PreStimSilence = 0
if pre_dur is None:
pre_dur = PreStimSilence
if occurrence is not None:
# Get values from specified occurrence and channel
extracted = val['resp'].extract_epoch(epoch)
r_vector = extracted[occurrence][c]
else:
r_vector = val['resp'].as_continuous()[0, :]
validbins = np.isfinite(r_vector)
r_vector = nems.utils.smooth(r_vector[validbins], 7)
r_vector = r_vector[3:-3]
# Convert bins to time (relative to start of epoch)
# TODO: want this to be absolute time relative to start of data?
time_vector = np.arange(0, len(r_vector)) / val['resp'].fs - PreStimSilence
# limit time range if specified
good_bins = (time_vector >= -pre_dur)
if dur is not None:
good_bins[time_vector > dur] = False
if occurrence is not None:
extracted = val['pred'].extract_epoch(epoch)
p_vector = extracted[occurrence][c]
else:
p_vector = val['pred'].as_continuous()
p_vector = p_vector[0, validbins]
times.append(time_vector[good_bins])
values.append(r_vector[good_bins] + i)
times.append(time_vector[good_bins])
values.append(p_vector[good_bins] + i)
r_test.append(ctx['modelspec'].meta['r_test'][0])
times_all = times
values_all = values
cc12 = np.corrcoef(values_all[0], values_all[1])[0, 1]
cc13 = np.corrcoef(values_all[0], values_all[2])[0, 1]
cc23 = np.corrcoef(values_all[1], values_all[2])[0, 1]
ccd23 = np.corrcoef(values_all[1]-values_all[0],
values_all[2]-values_all[0])[0, 1]
# ccd12 = np.corrcoef(values_all[0]-values_all[3],
# values_all[1]-values_all[3])[0, 1]
# ccd13 = np.corrcoef(values_all[0]-values_all[3],
# values_all[2]-values_all[3])[0, 1]
# ccd23 = np.corrcoef(values_all[1]-values_all[3],
# values_all[2]-values_all[3])[0, 1]
print("CC LN-GC: {:.3f} LN-STP: {:.3f} STP-GC: {:.3f}".format(
cc12,cc13,cc23))
# print("CCd LN-GC: {:.3f} LN-STP: {:.3f} STP-GC: {:.3f}".format(
# ccd12,ccd13,ccd23))
plt.figure()
ax = plt.subplot(2, 1, 1)
if occurrence is not None:
extracted = val['stim'].extract_epoch(epoch)
sg = extracted[occurrence]
else:
sg = val['stim'].as_continuous()
sg = sg[:, validbins]
sg = sg[:, good_bins]
nplt.plot_spectrogram(sg, val['resp'].fs, ax=ax,
title='{} Stim {}'.format(cellid, occurrence),
time_offset=pre_dur, cmap='gist_yarg')
ax = plt.subplot(2, 1, 2)
title = 'Preds LN {:.3f} GC {:.3f} STP {:.3f} /CC LN-GC: {:.3f} LN-STP: {:.3f} STP-GC: {:.3f} dSTP-dGC: {:.3f}'.format(
r_test[0],r_test[1],r_test[2],cc12,cc13,cc23,ccd23)
nplt.plot_timeseries(times_all, values_all, ax=ax, legend=legend,
title=title, colors=plot_colors)
plt.tight_layout()
return cc12, cc13, cc23, ccd23
Ya = rec_avg['resp'].as_continuous()
print('Trial-averaged stim/resp extracted to matrices Xa=({},{}) / Ya=({},{})'.
format(Xa.shape[0], Xa.shape[1], Ya.shape[0], Ya.shape[1]))
# here's how you get rasters aligned to each stimulus:
# define regex for stimulus epochs
epoch_regex = '^STIM_'
epochs_to_extract = ep.epoch_names_matching(rec.epochs, epoch_regex)
folded_resp = resp.extract_epochs(epochs_to_extract)
print('Response to each stimulus extracted to folded_resp dictionary')
# or just plot the PSTH for an example stimulus
raster = resp.extract_epoch(epoch)
psth = np.mean(raster, axis=0)
spec = stim.extract_epoch(epoch)[0, :, :]
plt.figure()
ax = plt.subplot(2, 1, 1)
nplt.plot_spectrogram(spec, fs=resp.fs, ax=ax, title=epoch)
ax = plt.subplot(2, 1, 2)
nplt.timeseries_from_vectors(psth,
fs=resp.fs,
ax=ax,
title="PSTH ({} cells, {} reps)".format(
raster.shape[1], raster.shape[0]))
plt.tight_layout()
ax = plt.subplot(2, 2, 1)
ax.imshow(sc, aspect='equal', cmap='gray')
mm = np.max(sc[sc < 1])
i, j = np.where(sc == mm)
ax = plt.subplot(2, 1, 2)
nplt.timeseries_from_vectors(psth,
fs=resp.fs,
ax=ax,
title="PSTH ({} cells, {} reps)".format(
raster.shape[1], raster.shape[0]))
fig = plt.figure()
ax = plt.subplot(3, 1, 1)
nplt.plot_spectrogram(spec, fs=resp.fs, ax=ax, title=epoch, cmap='gray_r')
ax = plt.subplot(3, 1, 2)
tt = np.arange(psth.shape[1]) / resp.fs - 2
ax.plot(tt, psth[25, :], color='black', lw=1)
ax.plot(tt, psth[27, :], color='gray', lw=1)
ax.set_title('sc={:.3f}'.format(sc[25, 27]))
ax.set_ylabel('Spikes/sec')
nplt.ax_remove_box(ax)
ax = plt.subplot(3, 1, 3)
ax.plot(tt, psth[25, :], color='black', lw=1)
ax.plot(tt, psth[32, :], color='gray', lw=1)
ax.set_title('sc={:.3f}'.format(sc[25, 32]))
def compare_model_preds(cellid,
batch,
modelname1,
modelname2,
max_pre=0.25,
max_dur=1.0):
"""
compare prediction accuracy of two models on validation stimuli
borrows a lot of functionality from nplt.quickplot()
"""
xf1, ctx1 = get_model_preds(cellid, batch, modelname1)
xf2, ctx2 = get_model_preds(cellid, batch, modelname2)
rec = ctx1['rec']
val1 = ctx1['val']
val2 = ctx2['val']
stim = rec['stim'].rasterize()
resp = rec['resp'].rasterize()
pred1 = val1['pred']
pred2 = val2['pred']
fs = resp.fs
d = resp.get_epoch_bounds('PreStimSilence')
PreStimSilence = np.mean(np.diff(d))
d = resp.get_epoch_bounds('PostStimSilence')
PostStimSilence = np.mean(np.diff(d))
epoch_regex = "^STIM_"
stim_epochs = ep.epoch_names_matching(resp.epochs, epoch_regex)
r = resp.as_matrix(stim_epochs)
s = stim.as_matrix(stim_epochs)
repcount = np.sum(np.isfinite(r[:, :, 0, 0]), axis=1)
max_rep_id, = np.where(repcount == np.max(repcount))
# keep a max of two stimuli
stim_ids = max_rep_id[:2]
stim_count = len(stim_ids)
# print(max_rep_id)
# stim_i=max_rep_id[-1]
# print("Max rep stim={} ({})".format(stim_i, stim_epochs[stim_i]))
p1 = pred1.as_matrix(stim_epochs)
p2 = pred2.as_matrix(stim_epochs)
ms1 = ctx1['modelspec']
ms2 = ctx2['modelspec']
r_test1 = ms1.meta['r_test'][0]
r_test2 = ms2.meta['r_test'][0]
fh = plt.figure(figsize=(16, 6))
# model 1 modules
ax = plt.subplot(5, 4, 1)
nplt.strf_timeseries(ms1,
ax=ax,
clim=None,
show_factorized=True,
title="{}/{} rtest={:.3f}".format(
cellid, modelname1, r_test1),
fs=resp.fs)
nplt.ax_remove_box(ax)
ax = plt.subplot(5, 4, 3)
nplt.strf_timeseries(ms2,
ax=ax,
clim=None,
show_factorized=True,
title="{}/{} rtest={:.3f}".format(
cellid, modelname2, r_test2),
fs=resp.fs)
nplt.ax_remove_box(ax)
if find_module('stp', ms1):
ax = plt.subplot(5, 4, 5)
nplt.before_and_after_stp(ms1,
sig_name='pred',
ax=ax,
title='',
channels=0,
xlabel='Time (s)',
ylabel='STP',
fs=resp.fs)
nplt.ax_remove_box(ax)
nlidx = find_module('double_exponential', ms1, find_all_matches=True)
if len(nlidx):
nlidx = nlidx[-1]
fn1, fn2 = nplt.before_and_after_scatter(rec,
ms1,
nlidx,
smoothing_bins=200,
mod_name='double_exponential')
ax = plt.subplot(5, 4, 6)
fn1(ax=ax)
nplt.ax_remove_box(ax)
# model 2 modules
wcidx = find_module('weight_channels', ms2)
if wcidx:
ax = plt.subplot(5, 4, 4)
coefs = ms2[wcidx]['phi']['coefficients']
plt.imshow(coefs,
clim=np.array([-1, 1]) * np.max(np.abs(coefs)),
cmap='bwr')
plt.xlabel('in')
plt.ylabel('out')
plt.colorbar()
nplt.ax_remove_box(ax)
if find_module('stp', ms2):
ax = plt.subplot(5, 4, 7)
nplt.before_and_after_stp(ms2,
sig_name='pred',
ax=ax,
title='',
channels=0,
xlabel='Time (s)',
ylabel='STP',
fs=resp.fs)
nplt.ax_remove_box(ax)
nlidx = find_module('double_exponential', ms2, find_all_matches=True)
if len(nlidx):
nlidx = nlidx[-1]
fn1, fn2 = nplt.before_and_after_scatter(rec,
ms2,
nlidx,
smoothing_bins=200,
mod_name='double_exponential')
ax = plt.subplot(5, 4, 8)
fn1(ax=ax)
nplt.ax_remove_box(ax)
max_bins = int((PreStimSilence + max_dur) * fs)
pre_cut_bins = int((PreStimSilence - max_pre) * fs)
if pre_cut_bins < 0:
pre_cut_bins = 0
else:
PreStimSilence = max_pre
for i, stim_i in enumerate(stim_ids):
ax = plt.subplot(5, 2, 5 + i)
if s.shape[2] <= 2:
nplt.timeseries_from_vectors(
[
s[stim_i, 0, 0, pre_cut_bins:max_bins],
s[max_rep_id[-1], 0, 1, pre_cut_bins:max_bins]
],
fs=stim.fs,
time_offset=PreStimSilence,
ax=ax,
title="{}/{} rfit={:.3f}/{:.3f}".format(
cellid, stim_epochs[stim_i], r_test1, r_test2))
else:
nplt.plot_spectrogram(s[stim_i, 0, :, pre_cut_bins:max_bins],
fs=stim.fs,
time_offset=PreStimSilence,
ax=ax,
title="{}/{} rfit={:.3f}/{:.3f}".format(
cellid, stim_epochs[stim_i], r_test1,
r_test2))
ax.get_xaxis().set_visible(False)
nplt.ax_remove_box(ax)
ax = plt.subplot(5, 2, 7 + i)
_r = r[stim_i, :, 0, pre_cut_bins:max_bins]
t = np.arange(_r.shape[-1]) / resp.fs - PreStimSilence - 0.5 / resp.fs
nplt.raster(t, _r)
ax.get_xaxis().set_visible(False)
ax = plt.subplot(5, 2, 9 + i)
nplt.timeseries_from_vectors([
np.nanmean(_r, axis=0), p1[stim_i, 0, 0, pre_cut_bins:max_bins],
p2[stim_i, 0, 0, pre_cut_bins:max_bins]
],
fs=resp.fs,
time_offset=PreStimSilence,
ax=ax)
nplt.ax_remove_box(ax)
plt.tight_layout()
return fh, ctx1, ctx2
for i in [4, 5, 7, 13]:
epoch = epochs_to_extract[i]
# or just plot the PSTH for an example stimulus
raster = resp.extract_epoch(epoch)
psth = np.mean(raster, axis=0)
spec = stim.extract_epoch(epoch)[0, :, :]
norm_psth = psth - np.mean(psth, axis=1, keepdims=True)
norm_psth /= np.std(psth, axis=1, keepdims=True)
f, ax = plt.subplots(3, 1, sharex=True)
PreStimSilence = 0.25
nplt.plot_spectrogram(spec,
fs=resp.fs,
ax=ax[0],
time_offset=PreStimSilence,
title=epoch,
cmap='gray_r')
nplt.plot_spectrogram(psth,
fs=resp.fs,
ax=ax[1],
time_offset=PreStimSilence,
cmap='gray_r')
PreStimSilence = 0.25
tt = np.arange(psth.shape[1]) / resp.fs - PreStimSilence
rasterfs = resp.fs
ax[2].plot(tt,
psth[25, :] * rasterfs,