def update_mu(self, z, wik, lr_mu, tau_mu, g_index=-1, max_iter=150):
N, D, M = z.shape + (wik.shape[-1], )
# if too much gaussian we compute mean for each gaussian separately (To avoid too large memory)
if (M > 40):
for g_index in range(M // 40 + (1 if (M % 40 != 0) else 0)):
from_ = g_index * 40
to_ = min((g_index + 1) * 40, M)
zz = z.unsqueeze(1).expand(N, to_ - from_, D)
self._mu[from_:to_] = pa.barycenter(zz,
wik[:, from_:to_],
lr_mu,
tau_mu,
max_iter=max_iter,
verbose=True,
normed=True).squeeze()
else:
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 _fast_maximisation(self, x, indexes, batch_size=10):
N, D = x.size(0), x.size(-1)
start_time = time.time()
centroids = x.new(self._n_c, x.size(-1))
barycenter_time = 0
mask_matrix = torch.zeros(x.size(0), self._n_c).to(x.device)
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)
indexes[random.randint(0, len(x) - 1)] = i
mask_matrix[indexes == i, i] = 1
nb_batch = (self._n_c // batch_size) + (1 if (
(self._n_c % batch_size) != 0) else 0)
for i in range(nb_batch):
start_index = i * batch_size
end_index = min((i + 1) * batch_size, self._n_c)
weight = mask_matrix[:, start_index:end_index]
barycenter_start_time = time.time()
centroids[start_index:end_index] = pa.barycenter(
x.unsqueeze(1).expand(N, end_index - start_index, D),
wik=weight,
normed=True,
lr=5e-3,
tau=1e-3,
verbose=self.verbose)
barycenter_end_time = time.time()
barycenter_time += (barycenter_end_time - barycenter_start_time)
end_time = time.time()
if (self.verbose):
print("Fast Maximisation Time ", end_time - start_time)
print("Cumulate barycenter Time ", barycenter_time)
return centroids
def _maximisation(self, x, indexes):
start_time = time.time()
centroids = x.new(self._n_c, x.size(-1))
barycenter_time = 0
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)
barycenter_start_time = time.time()
centroids[i] = pa.barycenter(lx, normed=True)
barycenter_end_time = time.time()
barycenter_time += (barycenter_end_time - barycenter_start_time)
end_time = time.time()
if (self.verbose):
print("Maximisation Time ", end_time - start_time)
print("Cumulate barycenter Time ", barycenter_time)
return centroids
loss = tree_embedding_criterion(pe_x, pe_y, z=ne,
manifold=manifold).sum()
tloss += loss.item()
loss.backward()
optimizer.step()
print('Loss value for iteration ', i, ' is ', tloss)
from rcome.function_tools import poincare_alg as pa
weigths = torch.Tensor([[1 if (y in dataset.Y[i]) else 0 for y in range(13)]
for i in range(len(X))]).cuda()
barycenters = []
for i in range(13):
barycenters.append(
pa.barycenter(model.weight.data, weigths[:, i], verbose=True).cpu())
plot_poincare_disc_embeddings(model.weight.data.cpu().numpy(),
labels=dataset.Y,
centroids=torch.cat(barycenters),
save_folder="LOG/mean",
file_name="LFR_hierachical.png")
print(barycenters)
barycenters = []
for i in range(13):
barycenters.append(
model.weight.data[weigths[:, i] == 1].mean(0).cpu().unsqueeze(0))
print(barycenters)
plot_poincare_disc_embeddings(model.weight.data.cpu().numpy(),
labels=dataset.Y,
import torch
import os
from rcome.visualisation_tools.plot_tools import plot_poincare_disc_embeddings
from rcome.function_tools import poincare_alg as pa
from matplotlib import pyplot as plt
data_point = (torch.randn(50, 2) / 5 + 0.6)
print(data_point)
norms = data_point.norm(2, -1)
data_point[norms >= 1.] = torch.einsum('ij, i -> ij', data_point[norms >= 1.],
1 / (norms[norms >= 1.] + 1e-5))
data_point.norm(2, -1)
barycenter_hyperbolic = pa.barycenter(data_point, verbose=True)
print(barycenter_hyperbolic)
plot_poincare_disc_embeddings(data_point.numpy(),
close=False,
save_folder="LOG/mean",
file_name="example_mean_hyperbolic.png")
barycenter_euclidean = data_point.mean(0, keepdim=True)
plt.scatter(barycenter_hyperbolic[:, 0],
barycenter_hyperbolic[:, 1],
marker='D',
s=300.,
c='red',
label="Hyperbolic")
plt.scatter(barycenter_euclidean[:, 0],