Add an example to visualize the evolutuion of fitted parameters in bounded parameter space #43
Conversation
|
This looks very nice, thank you! I will wait with the merge until I have the time to do a final thorough review in ~2 weeks (don't worry that this will be after the GSoC deadline). |
|
Will still do a bit of minor nit-picking later, but here's a thing I just encountered: while running the example, I get the following error (multiple times) during the 3D plotting: Traceback (most recent call last):
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/matplotlib/backend_bases.py", line 1187, in _on_timer
ret = func(*args, **kwargs)
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/matplotlib/animation.py", line 1449, in _step
still_going = Animation._step(self, *args)
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/matplotlib/animation.py", line 1169, in _step
self._draw_next_frame(framedata, self._blit)
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/matplotlib/animation.py", line 1188, in _draw_next_frame
self._draw_frame(framedata)
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/matplotlib/animation.py", line 1766, in _draw_frame
self._drawn_artists = self._func(framedata, *self._args)
File "/home/marcel/programming/brian2modelfitting/examples/hh_nevergrad_errorevolution.py", line 153, in animate
'r*', markersize=8, label='best params')
File "/home/marcel/anaconda3/envs/brian2modelfitting/lib/python3.7/site-packages/mpl_toolkits/mplot3d/axes3d.py", line 1421, in plot
zs = np.broadcast_to(zs, len(xs))
TypeError: len() of unsized objectIn case it matters, this is with matplotlib 3.2.1. Any idea what could cause this? |
|
The problem seems to be related to the version of (brian) alk@alk:~/github/brian2modelfitting/examples$ conda list | grep matplotlib
matplotlib 3.3.4 py38h06a4308_0
matplotlib-base 3.3.4 py38h62a2d02_0The mentioned problem is related to how 3-D plotting is performed in ax1.plot3D(res['g_k'], res['g_na'], res['g_l'], ...)This can be solved easily by inserting the data to a list: ax1.plot3D([res['g_k']], [res['g_na']], [res['g_l']], ...)The error occurred multiple times because the animation was repeated until it was stopped manually. Should I commit these changes? |
|
Ah, thanks for looking into this. I think it makes sense to have one element lists in the |
| from brian2 import * | ||
| from brian2modelfitting import * | ||
| import pandas as pd | ||
| from mpl_toolkits.mplot3d import Axes3D |
There was a problem hiding this comment.
This Axes3D import does not seem to be used.
|
|
||
| def callback(params, errors, best_params, best_error, index): | ||
| """Custom callback""" | ||
| print(f'[round {index + 1}]\t{np.min(errors)}') |
There was a problem hiding this comment.
Really minor thing, but maybe add an "Best error: " before the error to make the output a bit more meaningful.
| g_k=[6.e-07 * siemens, 6.e-05 * siemens]) | ||
|
|
||
| # visualization of best fitted traces | ||
| start_scope() |
There was a problem hiding this comment.
The start_scope shouldn't be needed here.
| # visualization of errors and parameters evolving over time | ||
| full_output = fitter.results(format='dataframe', use_units=False) | ||
| g_k = full_output['g_k'].to_numpy() | ||
| g_na = full_output['g_na'].to_numpy() | ||
| g_l = full_output['g_l'].to_numpy() | ||
| error = full_output['error'].to_numpy() |
There was a problem hiding this comment.
Yet another minor thing: the code might be easier to follow if it did not use the dataframe format here (that gets converted into numpy arrays right away), but instead format='dict'.
| ax1.set_xlabel('$g_K$ [S]') | ||
| ax1.set_ylabel('$g_{Na}$ [S]') | ||
| ax1.set_zlabel('$g_l$ [S]') | ||
| ax2.set_xlabel('round') | ||
| ax2.set_ylabel('error') |
There was a problem hiding this comment.
It's a bit of a matter of taste, but I prefer the slightly more compact set syntax, e.g.
ax1.set(xlabel='$g_K$ [S]', ylabel='$g_{Na}$ [S]', zlabel='$g_l$ [S]')| ax2.grid() | ||
|
|
||
|
|
||
| def init(): |
There was a problem hiding this comment.
I think the init and animate function could use a short comment for those not familiar with FuncAnimation.
Simple example that shows the error (and parameters) evolution in time by using the
nevergradoptimizer.Voltage traces are fitted using the default approach with the sample size of 40 for 25 rounds of fitting.
The first figure shows fitted traces and the input current that drives each neuron.
The second figure (animation) shows parameter population per each round of fitting (frame by frame), where the highlighted parameters are the one with the smallest overall error.
There is no additional dependencies used for this example other than
matplotlibwhich is already used bybrian2by default.Additionally, custom callback is also added in the
fitter.fit(...)method which is very similar to the one that can be found in the documentation under the Advanced Features section: Callback Function subsection.