elemwise.py 文件源码

def R_op(self, inputs, eval_points):
        outs = self(*inputs, **dict(return_list=True))
        rval = [None for x in outs]
        # For each output
        for idx, out in enumerate(outs):
            # make such that _bgrads computes only the gradients of the
            # current output on the inputs ( and not all outputs)
            ograds = [x.zeros_like() for x in outs]
            ograds[idx] = theano.tensor.ones_like(out)

            bgrads = self._bgrad(inputs, ograds)
            rop_out = None

            for jdx, (inp, eval_point) in enumerate(izip(inputs,
                                                    eval_points)):
                # if None, then we can just ignore this branch ..
                # what we do is to assume that for any non-differentiable
                # branch, the gradient is actually 0, which I think is not
                # the right thing to do .. have to talk to Ian and James
                # about it

                if bgrads[jdx] is None or \
                        isinstance(bgrads[jdx].type, DisconnectedType):
                    pass
                elif eval_point is not None:
                    if rop_out is None:
                        rop_out = bgrads[jdx] * eval_point
                    else:
                        rop_out = rop_out + bgrads[jdx] * eval_point

            rval[idx] = rop_out

        return rval