nnet.py 文件源码

def grad(self, inp, grads):
        dy, sm, y_idx = inp
        g_dx, = grads
        # TODO: currently we do not compute the gradient w.r.t. dy, because
        # advanced indexing is not working yet. When it works, do it to avoid
        # potentially misleading behavior in gradient computations! (although
        # typically we should not need the gradient w.r.t. dy).
        y_idx_range = tensor.arange(y_idx.shape[0])
        g_dy = tensor.sum(
            g_dx * subtensor.AdvancedIncSubtensor()(
                sm, tensor.fill(dy, -1), y_idx_range, y_idx), axis=1)
        g_sm = dy.dimshuffle(0, 'x') * g_dx
        g_y_idx = grad_not_implemented(self, 2, y_idx)
        return [g_dy, g_sm, g_y_idx]