def fit(self, x_train, y_train):
self._train_le(y_train)
self.n_samples = len(x_train)
self.sub_basis = pca_for_sets(
x_train, self.n_sdim, self.p_norm
).contiguous()
self._fit(x_train, y_train)
gpu_free()
def predict_proba(self, x_pred):
sub_basis = pca_for_sets(x_pred, self.n_sdim, self.p_norm)
sim, _, _ = subspacea_subs(self.sub_basis, sub_basis, self.metric_mat,
False)
sim = sim.cpu().numpy()
sim = self.lda.transform(sim.T)
sim = l2dist(self.dic, sim)
pred = [
1 / (1e-10 + np.min(sim[self.labels == i, :], axis=0))
for i in range(self.n_classes)
]
return np.asarray(pred)
def _fit(self, x_train, y_train):
cls_data = [
torch.cat(
[x_train[j]
for j in range(len(y_train)) if y_train[j] == i], 1)
for i in set(y_train)
]
sub_basis = (pca_for_sets(cls_data, self.n_sdim,
self.p_norm).contiguous().permute((2, 0, 1)))
gram_mat = (sub_basis @ sub_basis.permute((0, 2, 1))).sum(0)
full_dim = torch.matrix_rank(gram_mat)
_, eig_vec = torch.symeig(gram_mat, eigenvectors=True)
eig_vec = eig_vec.flip(1)[:, self.n_reducedim:full_dim]
self.metric_mat = eig_vec @ eig_vec.T
self.dic = ortha_subs(self.sub_basis, self.metric_mat)
def predict_proba(self, x_pred, ret_kmat=False):
if self.reddim_mat is not None:
if self.reddim_mat.shape[0] != self.reddim_mat.shape[1]:
x_pred = [self.reddim_mat @ i for i in x_pred]
sub_basis = pca_for_sets(x_pred, self.n_sdim, self.p_norm)
sub_basis = ortha_subs(sub_basis, self.metric_mat)
sim, _, _ = subspacea_subs(self.dic, sub_basis, self.metric_mat, False)
sim = sim.cpu().numpy()
pred = [
np.max(sim[self.labels == i, :], axis=0)
for i in range(self.n_classes)
]
if ret_kmat:
return np.asarray(pred), sim
return np.asarray(pred)