def reduce_samples(self, X, number_to_keep):
""" A method to compute the variance to the mean of samples for every sub-region
and select the sub-regions with the lowest variance.
The method doesn't consider the possibility that the number of sub-regions
vary from image to image since it doesn't make sense to reduce the number of
sub-regions in this case.
Inputs:
* X = a 3-D numpy array corresponding to machine learning
features. The first dimension is linked to the image where the
feature was computed, the second dimension is linked to the
sub-regions to compute the features.
* number_to_keep = number of samples to keep.
Parameters:
* parallel = a boolean to activate parallel computation or not.
* n_jobs = number of jobs to create for parallel computation.
Outputs:
* indexes = indexes corresponding to the data to keep"""
# Basic check
if number_to_keep >= X.shape[1]:
logging.warning('The number of sub-region to keep is greater or equal ' + \
'to the number of sub-regions.')
return X
# Compute variances
if self.parallel:
variances_list = \
Parallel(n_jobs=self.n_jobs)(delayed(compute_variance_one_region)(X[:,region_index,:],
self.distance_function, self.distance_args,
self.mean_function, self.mean_args)
for region_index in range(X.shape[1]))
else:
variances_list = []
for region_index in trange(X.shape[1]):
variances_list.append(
compute_variance_one_region(X[:, region_index, :],
self.distance_function,
self.distance_args,
self.mean_function,
self.mean_args))
# Making sure to discard nan values if it happens by assigning them to np.inf
variances_list = np.array(variances_list)
variances_list[np.isnan(variances_list)] = np.inf
# Sorting variances and keeping only the number_to_keep lowest features
indexes = variances_list.argsort()
return indexes[:number_to_keep]
