def _optimization_method(self, p, d_p, lr):
with torch.no_grad():
gradient = d_p
if (self.tau is None or self.v is None):
self.tau = torch.zeros(gradient.size()).to(p.device)
self.v = torch.zeros(gradient.size()).to(p.device)
# we update only weigth that moved
mask = (gradient.norm(2, -1) != 0)
if (mask.sum() == 0):
return
gradient = gradient[mask]
tau = self.tau[mask]
v = self.v[mask]
m = self.beta_1 * tau + (1 - self.beta_1) * gradient
self.v[mask] =\
self.beta_2 * v + (1 - self.beta_2) * pf.norm(p[mask], gradient)
self.v[mask] = torch.cat(
(v.unsqueeze(0), self.v[mask].unsqueeze(0))).max(0)[0]
updated_weight = pf.exp(p[mask],
-lr * m / torch.sqrt(self.v[mask]))
self.tau[mask] =\
pf.parallel_transport(from_point=p[mask],
to_point=updated_weight,
vector=m)
p[mask] = updated_weight
def barycenter(z, wik=None, lr=1e-3, tau=5e-6, max_iter=100, distance=pf.distance, normed=False,
init_method="default", verbose=False):
with torch.no_grad():
if(wik is None):
wik = 1.
barycenter = z.mean(0, keepdim=True)
else:
wik = wik.unsqueeze(-1).expand_as(z)
if(init_method == "global_mean"):
print("Bad init selected")
barycenter = z.mean(0, keepdim=True)
else:
barycenter = (z*wik).sum(0, keepdim=True)/wik.sum(0)
if(len(z) == 1):
return z
iteration = 0
cvg = math.inf
while(cvg>tau and max_iter>iteration):
iteration+=1
if(type(wik) != float):
grad_tangent = 2 * pf.log(barycenter.expand_as(z), z) * wik
nan_values = (~(barycenter == barycenter))
if(torch.nonzero(nan_values.squeeze()).shape[0]>0):
print("\n\n A At least one barycenter is Nan : ")
print(pf.log(barycenter.expand_as(z), z).sum(0))
print("index of nan values ", nan_values.squeeze().nonzero())
quit()
# torch 1.3 minimum for this operation
print("index of nan values ", nan_values.squeeze().nonzero())
else:
grad_tangent = 2 * pf.log(barycenter.expand_as(z), z)
if(normed):
if(type(wik) != float):
grad_tangent /= wik.sum(0, keepdim=True).expand_as(wik)
else:
grad_tangent /= len(z)
cc_barycenter = pf.exp(barycenter, lr * grad_tangent.sum(0, keepdim=True))
nan_values = (~(cc_barycenter == cc_barycenter))
if(torch.nonzero(nan_values.squeeze()).shape[0]>0):
print("\n\n At least one barycenter is Nan exp update may contain 0: ")
print(grad_tangent.sum(0, keepdim=True))
quit()
# torch 1.3 minimum for this operation
cvg = distance(cc_barycenter, barycenter).max().item()
barycenter = cc_barycenter
if(cvg<=tau and verbose):
print("Frechet Mean converged in ", iteration, " iterations")
return barycenter
def _optimization_method(self, p, d_p, lr):
with torch.no_grad():
a = pf.exp(p, -lr * d_p)
if (((a.norm(2, -1)) >= self.eps).max() > 0):
if (not self.first_over):
print("New update out of the disc:", a.norm(2, -1).max())
self.first_over = True
mask = a[a.norm(2, -1) >= self.eps]
a[a.norm(2, -1) >= self.eps] = p[a.norm(2, -1) >= self.eps]
p.copy_(a)
def _optimization_method(self, p, d_p, lr):
p.copy_(pf.exp(p, lr * d_p))
def _optimization_method(self, p, d_p, lr):
p.copy_(
pf.add(pf.renorm_projection(p.data),
-lr * pf.exp(d_p.new(d_p.size()).zero_(), d_p)))