In [1]:
import sys
sys.path.insert(0, '..')
import time
import matplotlib.pyplot as plt
%matplotlib inline
In [2]:
def plot_durations(t_current, t_other):
print('Currently used version: {:.5f}s'.format(t_current))
print('Version of other package: {:.5f}s'.format(t_other))
print('=> {:.2f}%'.format(100 * (t_other - t_current) / t_current), 'better performance')
plt.bar(range(2), [t_current, t_other], color="black", tick_label=['Currently used', 'Other package'])
plt.ylabel('Duration in seconds')
In [3]:
from lib.datasets import PascalVOC
from lib.segmentation import slic
image = PascalVOC('../test_data').test.next_batch(3, shuffle=False)[0][2]
segmentation = slic(image, num_segments=800, compactness=30, max_iterations=10, sigma=0)
Speed comparison of the adjacency generation out of a segmentation between the currently used version and the version in skimage.
In [4]:
from lib.segmentation import segmentation_adjacency
from skimage.future.graph import RAG
t_current = time.process_time()
segmentation_adjacency(segmentation)
t_current = time.process_time() - t_current
t_other = time.process_time()
RAG(segmentation)
t_other = time.process_time() - t_other
plot_durations(t_current, t_other)
Speed comparison of the feature extraction of segments between the currently used version and the version in skimage.
In [5]:
from lib.segmentation import FormFeatureExtraction
from skimage.measure import regionprops
t_current = time.process_time()
FormFeatureExtraction(segmentation).get_features()
t_current = time.process_time() - t_current
t_other = time.process_time()
props = regionprops(segmentation + 1)
features = []
for prop in props:
prop['area']
prop['bbox']
prop['local_centroid']
prop['eccentricity']
prop['equivalent_diameter']
prop['extent']
prop['moments_hu']
prop['inertia_tensor']
prop['inertia_tensor_eigvals']
prop['major_axis_length']
prop['minor_axis_length']
prop['moments_central']
prop['moments_normalized']
prop['orientation']
t_other = time.process_time() - t_other
plot_durations(t_current, t_other)
Speed comparison of the coarsening algorithm between the currently used version and the version used in cnn_graph.
In [6]:
from lib.segmentation import segmentation_adjacency
from lib.graph import coarsen_adj
import os
import imp
import requests
def import_cdn(uri):
name = os.path.basename(uri).lower().rstrip('.py')
r = requests.get(uri)
r.raise_for_status()
codeobj = compile(r.content, uri, 'exec')
module = imp.new_module(name)
exec (codeobj, module.__dict__)
return module
# Load the cnn_graph coarsen script.
cnn_graph = import_cdn('https://raw.githubusercontent.com/mdeff/cnn_graph/master/lib/coarsening.py')
adj, points, mass = segmentation_adjacency(segmentation)
t_current = time.process_time()
coarsen_adj(adj, points, mass, levels=4)
t_current = time.process_time() - t_current
# Run the coarsen script without slow prints.
t_other = time.process_time()
graphs, parents = cnn_graph.metis(adj, levels=4)
perms = cnn_graph.compute_perm(parents)
for i, A in enumerate(graphs):
M, M = A.shape
A = A.tocoo()
A.setdiag(0)
if i < 4:
A = cnn_graph.perm_adjacency(A, perms[i])
A = A.tocsr()
A.eliminate_zeros()
graphs[i] = A
t_other = time.process_time() - t_other
plot_durations(t_current, t_other)