gfilt.py 文件源码

def grad(self, inputs, ograds):
        ref, values, ref_dim, val_dim = inputs[:4]
        hash_struct = inputs[4:]
        ograd = ograds[0]

        ref_dim = get_scalar_constant_value(ref_dim)
        val_dim = get_scalar_constant_value(val_dim)

        def _conv(x):
            return GaussianFilter()(ref, x, ref_dim, val_dim, *hash_struct)

        # Since the kernels are separable and symmetric, the gradient w.r.t.
        # input is just the same filtering applied to the output grads.
        grad_i = _conv(ograd)

        def _gradr(r_i, vals, og, *args):
            return (og * (_conv(vals*r_i) - r_i*_conv(vals)) +
                    vals * (_conv(og*r_i) - r_i*_conv(og)))

        grad_r, _ = theano.scan(fn=_gradr, sequences=[ref],
                                non_sequences=[values, ograd] + hash_struct,
                                outputs_info=None)

        grad_r = grad_r.sum(axis=1, acc_dtype="float32")

        grads = [DisconnectedType()() for i in range(len(inputs))]
        grads[0] = grad_r
        grads[1] = grad_i
        return grads