def scale_to_sargent(candidate):
"""
Scale fit values to match peak AMPA reported in Sargent2005.
(unblocked, and phi=1!)
"""
sargent_peak = 0.63 # (nS)
timestep = 0.01
timepoints = np.arange(0, 300, timestep)
pulse_times = np.array([10.])
single_waveform_length = timepoints.shape[0]
signal = synthetic_conductance_signal(timepoints,
pulse_times,
single_waveform_length,
timestep,
0.,
*candidate)
scaling_factor = sargent_peak/signal.max()
scaled_signal = scaling_factor * signal
scaled_candidate = candidate[:]
for k in [1,2]:
# scale 'amplitude' parameters
scaled_candidate[k] *= scaling_factor
print(scaled_candidate)
# rounded scaled candidate should be [0.8647, 17.00, 2.645, 13.52,
# 121.9, 0.0322, 236.1, 6.394]
fig, ax = plt.subplots()
ax.plot(timepoints, signal, label="fit to Rothman data")
ax.plot(timepoints, scaled_signal, label="scaled to peak value by Sargent (unblocked, phi=1)")
ax.legend(loc="best")
fig.suptitle("parameters {0}".format(scaled_candidate))
plt.show()
def plot_single_comparison(filename, candidate, pulse_n, timepoints, trace, timestep=0.025):
pulse_times = problem.exp_pulses[pulse_n]
single_waveform_length = problem.single_waveform_lengths[8]
signal = synthetic_conductance_signal(timepoints,
pulse_times,
single_waveform_length,
timestep,
0.,
*candidate)
fig, ax = plt.subplots()
# plot trace we are comparing
ax.plot(timepoints, trace, linewidth=2, color='k')
# plot model trace
ax.plot(timepoints, signal, linewidth=2, color="r")
# plot spike raster
displacement = 0
ax.scatter(pulse_times,
np.zeros_like(pulse_times)-displacement,
marker="o",
color="k")
ax.set_xlabel('time (ms)')
ax.set_ylabel('amplitude (a.u.)')
ax.xaxis.set_ticks_position('bottom')
ax.yaxis.set_ticks_position('left')
ax.spines['top'].set_color('none')
ax.spines['right'].set_color('none')
plt.savefig(filename,
dpi=100,
size=(5,10))
plt.show()
def plot_optimisation_results(candidate, fitness, max_evaluations, pop_size):
"""
Plot a comparison of the model, the experimental data and the
original model by Rothman (Schwartz2012) across all experimental
traces. The figure gets saved to disk in the current folder.
"""
fig, ax = plt.subplots(nrows=8, ncols=4, figsize=(160,80), dpi=500)
rothman_fitness = 0
java_fit_time_points = []
java_fit_signals = []
with h5py.File("NMDA_jason_fit_traces.hdf5") as java_fit_data_repo:
for freq in problem.frequencies:
for prot in range(problem.n_protocols):
data_group = java_fit_data_repo['/{0}/{1}'.format(freq, prot)]
# jason's modeled traces have their first peak
# normalised to 1 and they don't have any offset, so
# we must add 1ms to the time dimension and
# 'denormalise' them by scaling them by the value he
# measured as the experimental NMDA peak conductance
# (0.18nS)
java_fit_time_points.append(np.array(data_group['average_waveform'][:-1,0]) + 1.)
java_fit_signals.append(np.array(data_group['average_waveform'][:-1,1]) * 0.18)
for k, ep in enumerate(problem.exp_pulses):
timepoints = problem.exp_data[k][:,0]
signal = synthetic_conductance_signal(timepoints,
ep,
problem.single_waveform_lengths[k],
problem.timestep_sizes[k],
1.,
*candidate)
# we use the trace we compute in Python to calculate the
# fitness of Jason's model, whereas the plotted trace is taken
# straight from the synthetic data he generated in Java. The
# two coincide anyway, and this is just to avoid headaches
# with time axis misalignments.
rothman_signal_python = rothman2012_NMDA_signal(timepoints,
ep,
problem.single_waveform_lengths[k],
problem.timestep_sizes[k])
rothman_signal_java = java_fit_signals[k]
rothman_fitness += np.linalg.norm(rothman_signal_python - problem.exp_data[k][:,1])/np.sqrt(timepoints.shape[0])
ax.flat[k].plot(timepoints, problem.exp_data[k][:,1], color='k', linewidth=3)
ax.flat[k].scatter(ep, np.zeros(shape=ep.shape)-0.1, color='k')
ax.flat[k].plot(java_fit_time_points[k], rothman_signal_java, linewidth=1, color='r')
ax.flat[k].plot(timepoints, signal, linewidth=1, color='g')
rothman_fitness /= len(problem.exp_pulses)
fig.suptitle('parameters: {0}\n fitness: {1} max_evaluations: {2} pop_size: {3}\nRothman2012 fitness: {4}'.format(candidate, fitness, max_evaluations, pop_size, rothman_fitness))
plt.savefig('Rothman_NMDA_TM_fit_{0}.png'.format(time.time()))
def fitness_to_experiment(cs):
# fitness is defined as the average (across experimental datasets)
# of the L2 norm between the synthetic and the experimental signal
# divided by the square root of the number of time points, to
# avoid unfair weighing in favour of longer recordings (in other
# words, this should give equal weight to every time point in
# every recording).
distances = []
for k, ep in enumerate(problem.exp_pulses):
timepoints = problem.exp_data[k][:,0]
signal = synthetic_conductance_signal(timepoints,
ep,
problem.single_waveform_lengths[k],
problem.timestep_sizes[k],
1.,
*cs)
distances.append(np.linalg.norm(signal-problem.exp_data[k][:,1])/np.sqrt(timepoints.shape[0]))
return sum(distances)/len(distances)
