def forward()

in ppuda/deepnets1m/net.py [0:0]

• 39 lines of code
• 28 McCabe index (conditional complexity)

    def forward(self, s0, s1, drop_path_prob=0):

        s0 = None if (s0 is None or _is_none(self.preprocess0)) else self.preprocess0(s0)
        s1 = None if (s1 is None or _is_none(self.preprocess1)) else self.preprocess1(s1)

        states = [s0, s1]
        for i in range(self._steps):
            h1 = states[self._indices[2 * i]]
            h2 = states[self._indices[2 * i + 1]]
            op1 = self._ops[2 * i]
            op2 = self._ops[2 * i + 1]
            s = None

            if not (isinstance(op1, Zero) or _is_none(op1) or h1 is None):
                h1 = op1(h1)
                if self.training and drop_path_prob > 0 and not isinstance(op1, nn.Identity):
                    h1 = drop_path(h1, drop_path_prob)
                s = h1

            if not (isinstance(op2, Zero) or _is_none(op2) or h2 is None):
                h2 = op2(h2)
                if self.training and drop_path_prob > 0 and not isinstance(op2, nn.Identity):
                    h2 = drop_path(h2, drop_path_prob)
                try:
                    s = h2 if s is None else (h1 + h2)
                except:
                    print(h1.shape, h2.shape, self.genotype)
                    raise

            states.append(s)


        if sum([states[i] is None for i in self._concat]) > 0:
            # Replace None states with Zeros to match feature dimensionalities and enable forward pass
            assert self._has_none, self.genotype
            s_dummy = None
            for i in self._concat:
                if states[i] is not None:
                    s_dummy = states[i] * 0
                    break

            if s_dummy is None:
                return None
            else:
                for i in self._concat:
                    if states[i] is None:
                        states[i] = s_dummy

        return torch.cat([states[i] for i in self._concat], dim=1)