def make_comparison_figures():
gt_sorting, tested_sorting = generate_erroneous_sorting()
comp = sc.compare_sorter_to_ground_truth(gt_sorting, tested_sorting, gt_name=None, tested_name=None,
delta_time=0.4, sampling_frequency=None, min_accuracy=0.5, exhaustive_gt=True, match_mode='hungarian',
n_jobs=-1, bad_redundant_threshold=0.2, compute_labels=False, verbose=False)
print(comp.hungarian_match_12)
fig, ax = plt.subplots()
im = ax.matshow(comp.match_event_count, cmap='Greens')
ax.set_xticks(np.arange(0, comp.match_event_count.shape[1]))
ax.set_yticks(np.arange(0, comp.match_event_count.shape[0]))
ax.xaxis.tick_bottom()
ax.set_yticklabels(comp.match_event_count.index, fontsize=12)
ax.set_xticklabels(comp.match_event_count.columns, fontsize=12)
fig.colorbar(im)
fig.savefig('spikecomparison_match_count.png')
fig, ax = plt.subplots()
sw.plot_agreement_matrix(comp, ax=ax, ordered=False)
im = ax.get_images()[0]
fig.colorbar(im)
fig.savefig('spikecomparison_agreement_unordered.png')
fig, ax = plt.subplots()
sw.plot_agreement_matrix(comp, ax=ax)
im = ax.get_images()[0]
fig.colorbar(im)
fig.savefig('spikecomparison_agreement.png')
fig, ax = plt.subplots()
sw.plot_confusion_matrix(comp, ax=ax)
im = ax.get_images()[0]
fig.colorbar(im)
fig.savefig('spikecomparison_confusion.png')
plt.show()
# The final part of this tutorial deals with comparing spike sorting outputs.
# We can either (1) compare the spike sorting results with the ground-truth sorting :code:`sorting_true`, (2) compare
# the output of two (Klusta and Mountainsor4), or (3) compare the output of multiple sorters:
comp_gt_KL = sc.compare_sorter_to_ground_truth(gt_sorting=sorting_true,
tested_sorting=sorting_KL)
comp_KL_MS4 = sc.compare_two_sorters(sorting1=sorting_KL, sorting2=sorting_MS4)
comp_multi = sc.compare_multiple_sorters(
sorting_list=[sorting_MS4, sorting_KL], name_list=['klusta', 'ms4'])
##############################################################################
# When comparing with a ground-truth sorting extractor (1), you can get the sorting performance and plot a confusion
# matrix
comp_gt_KL.get_performance()
w_conf = sw.plot_confusion_matrix(comp_gt_KL)
##############################################################################
# When comparing two sorters (2), we can see the matching of units between sorters. For example, this is how to extract
# the unit ids of Mountainsort4 (sorting2) mapped to the units of Klusta (sorting1). Units which are not mapped has -1
# as unit id.
mapped_units = comp_KL_MS4.get_mapped_sorting1().get_mapped_unit_ids()
print('Klusta units:', sorting_KL.get_unit_ids())
print('Mapped Mountainsort4 units:', mapped_units)
##############################################################################
# When comparing multiple sorters (3), you can extract a :code:`SortingExtractor` object with units in agreement
# between sorters. You can also plot a graph showing how the units are matched between the sorters.
perf = cmp_gt_HS.get_performance() ############################################################################## # Lets use seaborn swarm plot fig1, ax1 = plt.subplots() perf2 = pd.melt(perf, var_name='measurement') ax1 = sns.swarmplot(data=perf2, x='measurement', y='value', ax=ax1) ax1.set_xticklabels(labels=ax1.get_xticklabels(), rotation=45) ############################################################################## # The confusion matrix is also a good summary of the score as it has # the same shape as agreement matrix, but it contains an extra column for FN # and an extra row for FP sw.plot_confusion_matrix(cmp_gt_HS) ############################################################################## # We can query the well and bad detected units. By default, the threshold # on accuracy is 0.95. cmp_gt_HS.get_well_detected_units() ############################################################################## cmp_gt_HS.get_false_positive_units() ############################################################################## cmp_gt_HS.get_redundant_units()
def test_confusion(self):
sw.plot_confusion_matrix(self._gt_comp, count_text=True)
sorting_out=hither.File())
#Aggregating the output of the sorters
sorting_MS4 = AutoSortingExtractor(result_MS4.outputs.sorting_out._path)
sorting_SP = AutoSortingExtractor(
result_spyKingCircus.outputs.sorting_out._path)
#Comparing each to ground truth-confusion matrix
comp_MATLAB = sc.compare_sorter_to_ground_truth(gtOutput,
sortingPipeline,
sampling_frequency=sampleRate,
delta_time=3,
match_score=0.5,
chance_score=0.1,
well_detected_score=0.1,
exhaustive_gt=True)
w_comp_MATLAB = sw.plot_confusion_matrix(comp_MATLAB, count_text=True)
plt.show()
comp_MS4 = sc.compare_sorter_to_ground_truth(gtOutput,
sorting_MS4,
sampling_frequency=sampleRate,
delta_time=3,
match_score=0.5,
chance_score=0.1,
well_detected_score=0.1,
exhaustive_gt=True)
w_comp_MS4 = sw.plot_confusion_matrix(comp_MS4, count_text=True)
plt.show()
comp_SP = sc.compare_sorter_to_ground_truth(gtOutput,
sorting_SP,
# Widgets using SortingComparison # --------------------------------- # # We can compare the spike sorting output to the ground-truth sorting :code:`sorting_true` using the # :code:`comparison` module. :code:`comp_MS4` and :code:`comp_KL` are :code:`SortingComparison` objects import spikeinterface.comparison as sc comp_MS4 = sc.compare_sorter_to_ground_truth(sorting_true, sorting_MS4) comp_KL = sc.compare_sorter_to_ground_truth(sorting_true, sorting_KL) ############################################################################## # plot_confusion_matrix() # ~~~~~~~~~~~~~~~~~~~~~~~~~~ w_comp_MS4 = sw.plot_confusion_matrix(comp_MS4, count_text=False) w_comp_KL = sw.plot_confusion_matrix(comp_KL, count_text=False) ############################################################################## # plot_agreement_matrix() # ~~~~~~~~~~~~~~~~~~~~~~~~~~ w_agr_MS4 = sw.plot_agreement_matrix(comp_MS4, count_text=False) ############################################################################## # plot_sorting_performance() # ~~~~~~~~~~~~~~~~~~~~~~~~~~ # # We can also plot a performance metric (e.g. accuracy, recall, precision) with respect to a quality metric, for # example signal-to-noise ratio. Quality metrics can be computed using the :code:`toolkit.validation` submodule
perf = cmp_gt_MS4.get_performance() ############################################################################## # Lets use seaborn swarm plot fig1, ax1 = plt.subplots() perf2 = pd.melt(perf, var_name='measurement') ax1 = sns.swarmplot(data=perf2, x='measurement', y='value', ax=ax1) ax1.set_xticklabels(labels=ax1.get_xticklabels(), rotation=45) ############################################################################## # The confusion matrix is also a good summary of the score as it has # the same shape as agreement matrix, but it contains an extra column for FN # and an extra row for FP sw.plot_confusion_matrix(cmp_gt_MS4) ############################################################################## # We can query the well and bad detected units. By default, the threshold # on accuracy is 0.95. cmp_gt_MS4.get_well_detected_units() ############################################################################## cmp_gt_MS4.get_false_positive_units() ############################################################################## cmp_gt_MS4.get_redundant_units()
# So you can acces finely to all individual results.
#
# Note that exhaustive_gt=True when you excatly how many
# units in ground truth (for synthetic datasets)
study.run_comparisons(exhaustive_gt=True)
for (rec_name, sorter_name), comp in study.comparisons.items():
print('*' * 10)
print(rec_name, sorter_name)
print(comp.count) # raw counting of tp/fp/...
comp.print_summary()
perf_unit = comp.get_performance(method='by_unit')
perf_avg = comp.get_performance(method='pooled_with_average')
m = comp.get_confusion_matrix()
w_comp = sw.plot_confusion_matrix(comp)
w_comp.ax.set_title(rec_name + ' - ' + sorter_name)
##############################################################################
# Collect synthetic dataframes and display
# -------------------------------------------------------------
#
# As shown previously, the performance is returned as a pandas dataframe.
# The :code:`aggregate_performances_table` function, gathers all the outputs in
# the study folder and merges them in a single dataframe.
dataframes = study.aggregate_dataframes()
##############################################################################
# Pandas dataframes can be nicely displayed as tables in the notebook.
