def indexed_forward(self, x, shift=None):
if self.positive:
positive(self.features)
self.grid.data = torch.clamp(self.grid.data, -1, 1)
N, c, w, h = x.size()
m = self.gauss_pyramid.scale_n + 1
feat = self.features.view(1, m * c, self.outdims)
if shift is None:
grid = self.grid.expand(N, self.outdims, 1, 2)
else:
grid = self.grid.expand(N, self.outdims, 1, 2) + shift[:, None, None, :]
pools = []
for xx in self.gauss_pyramid(x):
_, _, img_w, img_h = xx.size()
img_indexer = torch.tensor([(img_w -1) * img_h, img_h - 1]).type_as(grid)
(grid + 1) / 2
img_index = torch.round((grid + 1) / 2 * (img_shape - 1))
pools.append(F.grid_sample(xx, adj_grid))
y = torch.cat(pools, dim=1).squeeze(-1)
y = (y * feat).sum(1).view(N, self.outdims)
if self.bias is not None:
y = y + self.bias
return y
def all_forward(self, x, shift=None):
if self.positive:
positive(self.features)
self.grid.data = torch.clamp(self.grid.data, -1, 1)
N, c, w, h = x.size()
m = self.gauss_pyramid.scale_n + 1
feat = self.features.view(m * c, self.outdims, 1, 1).permute([1, 0, 2, 3]).contiguous()
stacked = torch.cat(self.gauss_pyramid(x), dim=1)
y = F.conv2d(stacked, feat, self.bias)
_, _, yw, yh = y.size()
return y
def simple_forward(self, x, shift=None):
if self.positive:
positive(self.features)
N, ch, h, w = x.size()
feat = self.features.view(1, -1, self.outdims)
feat = feat[:, 0:ch, :]
ctr_h, ctr_w = h // 2, w // 2
y = x[..., ctr_h, ctr_w].unsqueeze(-1).expand(N, ch, self.outdims)
y = (y * feat).sum(1).view(N, self.outdims)
if self.bias is not None:
y = y + self.bias
return y
def fixed_forward(self, x, shift=None):
if self.positive:
positive(self.features)
N, ch, h, w = x.size()
feat = self.features.view(1, -1, self.outdims)
feat_scale = feat.mean()
ctr_h, ctr_w = h // 2, w // 2
y = x[..., ctr_h, ctr_w].unsqueeze(-1).expand(N, ch, self.outdims)
y = (y * feat_scale).sum(1).view(N, self.outdims)
if self.bias is not None:
y = y + self.bias
return y
def disc_center_forward(self, x, shift=None):
if self.positive:
positive(self.features)
self.grid.data = torch.clamp(self.grid.data, -1, 1)
N, c, h, w = x.size()
m = self.gauss_pyramid.scale_n + 1
feat = self.features.view(1, m * c, self.outdims)
pools = []
for xx in self.gauss_pyramid(x):
N, ch, img_h, img_w = xx.size()
ctr_h, ctr_w = img_h // 2, img_w // 2
pools.append(xx[..., ctr_h, ctr_w].unsqueeze(-1).expand(N, ch, self.outdims))
y = torch.cat(pools, dim=1)
y = (y * feat).sum(1).view(N, self.outdims)
if self.bias is not None:
y = y + self.bias
return y