def prepare_network(self, run, epoch):
if self.model is None:
if self.network == 'dir':
model = DirectArch(miniXception_loader, self.net_input_shape, objective="malignancy", output_size=self.net_out_size,
normalize=self.net_normalize, pooling=self.net_pool)
elif self.network == 'siam':
model = SiamArch(miniXception_loader, self.net_input_shape, distance='l2', output_size=self.net_out_size, normalize=self.net_normalize,
pooling=self.net_pool)
elif self.network == 'dirR':
model = DirectArch(miniXception_loader, self.net_input_shape, objective="rating", output_size=self.net_out_size,
normalize=self.net_normalize, pooling=self.net_pool)
elif self.network == 'siamR':
model = SiamArch(miniXception_loader, self.net_input_shape, distance='l2', output_size=self.net_out_size, normalize=self.net_normalize,
pooling=self.net_pool, objective="rating")
elif self.network == 'trip':
model = TripArch(miniXception_loader, self.net_input_shape, distance='l2', output_size=self.net_out_size, normalize=self.net_normalize,
pooling=self.net_pool, categorize=self.categorize)
else:
assert (False)
else:
model = self.model
w = self.Weights(run=run, epoch=epoch)
assert (w is not None)
if self.network == 'dir':
embed_model = model.extract_core(weights=w, repool=False)
else:
embed_model = model.extract_core(weights=w)
return embed_model
H_lim, S_lim, Tau_lim, Conc_lim, Crst_lim = (1e6, 0), (1e6, 0), (1e6, 0), (1e6, 0), (1e6, 0)
for p, pooling in enumerate(pooling_options):
print('Evaluating {} pooling'.format(pooling))
h, s, tau, conc, crst = [], [], [], [], []
plot_data_filename = './Plots//Data/init_{}.p'.format(pooling)
try:
h, s, tau, conc, crst = pickle.load(open(plot_data_filename, 'br'))
print('loaded cached data for ' + pooling)
except:
start = timer()
for i in range(repeatitions):
print('\tRep # {}'.format(i))
model = DirectArch(miniXception_loader, input_shape, objective="malignancy", pooling=pooling,
output_size=out_size, normalize=True)
core = model.extract_core()
embed = core.predict(np.expand_dims(images, axis=-1), batch_size=32)
nbrs = NearestNeighbors(n_neighbors=N, algorithm='auto', metric=metric).fit(embed)
distances, indices = nbrs.kneighbors(embed)
distances, indices = distances[:, 1:], indices[:, 1:]
h += [index.calc_hubness(indices)[0]]
s += [index.calc_symmetry(indices)[0]]
tau += [index.kumar(distances, res=0.01)[0]]
conc += [index.concentration(distances)]
crst += [index.relative_contrast_imp(distances)]
pickle.dump((h, s, tau, conc, crst), open(plot_data_filename, 'bw'))
print('evaluated (and cached) {} in {:.1f} minutes '.format(pooling, (timer() - start) / 60))