def forward(self, x, centre=None, other=None, penultimate=False):
x = self.features(x)
x = F.interpolate(x, size=self.input_sz, mode="bilinear")
if not penultimate:
assert (centre is not None) and (other is not None)
patches1, patches2 = get_patches(x, centre, other, self.patch_side)
adjacency = self.isola_head(patches1, patches2)
x = torch.sigmoid(adjacency)
return x
def forward(self, x, centre=None, other=None, penultimate=False):
x = self.features(x)
x = F.interpolate(x, size=self.input_sz, mode="bilinear")
if not penultimate:
assert (centre is not None) and (other is not None)
patches1, patches2 = get_patches(x, centre, other, self.patch_side)
# predicted position distribution, no softmax - using
# torch.CrossEntropyLoss
# shape: bn, 9
x = self.doersch_head(patches1, patches2)
return x