def feature_extract(self, z, f1, f2=None, f3=None):
h = f1(z)
if f2 is not None and f3 is not None:
h_in = h
for f in f2:
h2 = f(h_in)
if self.model_name == "SRFlow-DA-D":
h2 = h2 + h_in
h_in = h2
if self.model_name == "SRFlow-DA-R":
h2 = h2 + h
h = f3(h2)
shift, scale = thops.split_feature(h, "cross")
scale = (torch.sigmoid(scale + 2.) + self.affine_eps)
return scale, shift
def split2d_prior(self, z, ft):
if ft is not None:
z = torch.cat([z, ft], dim=1)
h = self.conv(z)
return thops.split_feature(h, "cross")
def feature_extract_aff(self, z1, ft, f):
z = torch.cat([z1, ft], dim=1)
h = f(z)
shift, scale = thops.split_feature(h, "cross")
scale = (torch.sigmoid(scale + 2.) + self.affine_eps)
return scale, shift
def feature_extract(self, z, f):
h = f(z)
shift, scale = thops.split_feature(h, "cross")
scale = (torch.sigmoid(scale + 2.) + self.affine_eps)
return scale, shift