def compute_view_1(self, X):
""" compute network output of view 1 """
X = X.copy()
# # normalize data
# if self.prepare_view_1 is not None:
# X = self.prepare_view_1(X)
# return self.compute_v1_latent(X, self.dummy_in_v2)
dummy_in_v2 = np.repeat(self.dummy_in_v2, X.shape[0], axis=0)
return batch_compute2(X,
dummy_in_v2,
self.compute_v1_latent,
batch_size=min(10, X.shape[0]),
prepare1=self.prepare_view_1)
def compute_view_2(self, Z):
""" compute network output of view 2 """
Z = Z.copy()
# # normalize data
# if self.prepare_view_2 is not None:
# Z = self.prepare_view_2(Z)
# return self.compute_v2_latent(self.dummy_in_v1, Z)
dummy_in_v1 = np.repeat(self.dummy_in_v1, Z.shape[0], axis=0)
return batch_compute2(dummy_in_v1,
Z,
self.compute_v2_latent,
batch_size=min(10, Z.shape[0]),
prepare2=self.prepare_view_2)
l_v1latent, deterministic=True))
compute_v2_latent = theano.function(inputs=input_2,
outputs=lasagne.layers.get_output(
l_v2latent, deterministic=True))
print("Evaluating on test set...")
# compute output on test set
eval_set = 'test'
n_test = args.n_test if args.n_test is not None else data[eval_set].shape[0]
X1, X2 = data[eval_set][0:n_test]
print("Computing embedding ...")
lv1_latent = batch_compute2(X1,
X2,
compute_v1_latent,
np.min([100, n_test]),
prepare=model.prepare)
lv2_latent = batch_compute2(X1,
X2,
compute_v2_latent,
np.min([100, n_test]),
prepare=model.prepare)
if args.V2_to_V1:
lv1_latent, lv2_latent = flip_variables(lv1_latent, lv2_latent)
# reset n_test
n_test = lv1_latent.shape[0]
# evaluate retrieval result