def forward(self, x):
if (not self.evaluation):
# Create the veronese map of z
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
dim_veronese, BS = v_x.size()
M_inv_temp = self.create_M(v_x).cuda(1)
#TODO: Update Minv batch wise with sherman-morrisson techniques
M_inv = ((self.M_inv_copy.cuda(1) + M_inv_temp) * 0.5)
del M_inv_temp
torch.cuda.empty_cache()
x = torch.matmul(
torch.matmul(v_x.view(BS, 1, dim_veronese), M_inv),
v_x.view(BS, dim_veronese, 1))
self.M_inv_copy = M_inv.cpu().detach().clone()
else:
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
dim_veronese, BS = v_x.size()
x = torch.matmul(
torch.matmul(v_x.view(BS, 1, dim_veronese),
self.M_inv_copy.cuda(1)),
v_x.view(BS, dim_veronese, 1))
return x
def forward(self, x):
if (not self.has_M_inv and self.build_M):
# We want only veronese maps to build the moment matrix once we have good reconstruction (self.build_M = True) !
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
self.veroneses.append(v_x.cpu())
elif (self.has_M_inv):
# Create the veronese map of z
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
dim_veronese, BS = v_x.size()
x = torch.matmul(
torch.matmul(v_x.view(BS, 1, dim_veronese), self.M_inv),
v_x.view(BS, dim_veronese, 1))
return x
def forward(self, x):
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
# v_x is (dim_veronese, BS)
x = self.B(v_x)
return x
def forward(self, x):
npoints, dims = x.size()
v_x, _ = generate_veronese(x.view(dims, npoints), self.n)
# v_x is (dim_veronese, BS), transpose it to have the batch dim at the beginning
x = self.B(v_x.t_(), v_x)
return x