def _maximisation(self, x, indexes):
centroids = x.new(self._n_c, x.size(-1))
for i in range(self._n_c):
lx = x[indexes == i]
if (lx.shape[0] <= self._mec):
lx = x[random.randint(0, len(x) - 1)].unsqueeze(0)
centroids[i] = pa.barycenter(lx, normed=True)
return centroids
def update_mu(self, z, wik, lr_mu, tau_mu, g_index=-1, max_iter=50):
N, D, M = z.shape + (wik.shape[-1], )
# print(self._mu)
if (g_index > 0):
self._mu[g_index] = pa.barycenter(z,
wik[:, g_index],
lr_mu,
tau_mu,
max_iter=max_iter,
normed=True).squeeze()
else:
self._mu = pa.barycenter(z.unsqueeze(1).expand(N, M, D),
wik,
lr_mu,
tau_mu,
max_iter=max_iter,
normed=True).squeeze()
def accuracy_supervised(z, y, mu, nb_set=5, verbose=True):
n_example = len(z)
n_distrib = len(mu)
subset_index = torch.randperm(n_example)
nb_value = n_example // nb_set
I_CV = [
subset_index[nb_value * i:min(nb_value * (i + 1), n_example)]
for i in range(nb_set)
]
# print(I_CV)
acc_total = 0.
for i, test_index in enumerate(I_CV):
# create train dataset
train_index = torch.cat(
[subset for ci, subset in enumerate(I_CV) if (i != ci)], 0)
Z_train = z[train_index]
Y_train = torch.LongTensor([y[ic.item()] for ic in train_index])
#create test datase
Z_test = z[test_index]
Y_test = torch.LongTensor([y[ic.item()] for ic in test_index])
if (verbose):
print("Set " + str(i) + " :")
print("\t train size -> " + str(len(Z_train)))
print("\t test size -> " + str(len(Z_test)))
print("Obtaining centroids for each classes")
from function_tools import poincare_alg as pa
min_label = Y_train.min().item()
max_label = Y_train.max().item()
centroids = []
for i in range(n_distrib):
# print((Z_train[Y_train[:,0]== (min_label + i)]).size())
centroids.append(
pa.barycenter(Z_train[Y_train[:, 0] == (min_label + i)],
normed=False).tolist())
centroids = torch.Tensor(centroids).squeeze()
# predicting
Z_test_reshape = Z_test.unsqueeze(1).expand(Z_test.size(0), n_distrib,
Z_test.size(-1))
centroid_reshape = centroids.unsqueeze(0).expand_as(Z_test_reshape)
d2 = poincare_function.distance(Z_test_reshape, centroid_reshape)**2
predicted_labels = d2.min(-1)[1] + min_label
acc = (predicted_labels == Y_test.squeeze()).float().mean()
acc_total += acc.item()
print(acc)
return acc_total / (len(I_CV))