def compute_features(self, activations, agreement_threshold=0.3, analyze=False):
self.num_images = len(activations)
self.agreement_threshold = agreement_threshold
self.total_img_cxns = 0
self.image_activations = {}
for i, act in enumerate(activations):
asum = act.sum()
self.total_img_cxns += asum
self.image_activations[i] = (act, asum)
self.compute_feature_proposals()
print "Num proposals:", self.feature_idx
# print "Image features:", self.image_features
# print "Feature images:", self.feature_images
print "Compressions:\n", numpy.array(self.compressions)
# print "Fewer cxns:\n", numpy.array(self.feat_num_fewer_cxns)
num_sampling_attempts = 100
min_compression = 1.0
best_assignment = None
compressions = []
for i in xrange(num_sampling_attempts):
features, image_assignments, total_compression = self.sample_assignments()
if total_compression < min_compression:
min_compression = total_compression
best_assignment = image_assignments
best_features = features
compressions.append(total_compression)
histogram(compressions).show()
greedy_features, greedy_assignments, greedy_compression = self.get_greedy_assignment()
print "Best compression:", min_compression
print "Greedy compression:", greedy_compression
print best_assignment
coverages = []
for i, val in self.image_activations.items():
act, asum = val
feat_sum = greedy_features[greedy_assignments[i]].sum()
coverages.append(float(feat_sum) / asum)
print "Coverages:", coverages
return best_features, best_assignment
def randomly_sample_particle(self, num_samples):
""" Picks the best feature assignment from num_samples random attempts. """
num_sampling_attempts = 100
min_compression = 1.0
best_assignment = None
compressions = []
for i in xrange(num_sampling_attempts):
features, image_assignments, total_compression = \
self.randomly_condense_particles()
if total_compression < min_compression:
min_compression = total_compression
best_assignment = image_assignments
best_features = features
compressions.append(total_compression)
histogram(compressions).show()
