用户您好，欢迎安装使用mindspore。上述的问题已经收到，会尽快给您解答回复

用户您好，您的代码是否使用了反向过程，如果是的话那可能是反向传播不支持切片。您能提供一下您的代码我们再来看看具体是什么问题

代码内容比较多，不太好复现，这里给出了customwithlosscell，当调用self.stylegan2loss.accumulate_gradients时，计算如下方第二张图，return的值在construct里即为loss

class customwithlosscell(nn.cell):
    def __init__(self, g_mapping, g_synthesis, d):
        super(customwithlosscell, self).__init__()
        self.g_mapping = g_mapping
        self.g_synthesis = g_synthesis
        self.d = d
        self.stylegan2loss = stylegan2loss(self.g_mapping, self.g_synthesis, self.d, r1_gamma=5.0)

    # def construct(self, target_features, synth_features, reg_loss, regularize_noise_weight):
def construct(self, phase, real_img, pose, real_c, gen_z, gen_c, sync, gain):
        loss = self.stylegan2loss.accumulate_gradients(phase, real_img, pose, real_c, gen_z, gen_c, sync, gain)
        return loss

if do_gpl:
    batch_size = gen_z.shape[0] // self.pl_batch_shrink
    gen_ws = self.run_g_mapping(gen_z[:batch_size], gen_c[:batch_size], sync=sync)
    pl_grads = grad_op(net(self.g_synthesis))(gen_ws, pose[:batch_size])
    pl_lengths = sqrt(square(pl_grads).sum(2).mean(1))
    pl_mean = self.pl_mean   (pl_lengths.mean() - self.pl_mean) * self.pl_decay
    self.pl_mean = pl_mean.copy()
    pl_penalty = square(pl_lengths - pl_mean)
    loss_gpl = pl_penalty * self.pl_weight
    print('loss_gpl: ', loss_gpl.mean())
    return mul(loss_gpl.mean(), gain)

是的用户就是说grad_od来获取梯度过程

然后由于你给出的代码片段太少我们不好进行debug，能否麻烦你在gitee上创建一个仓库然后存放所有代码或者直接通过附件形式提供代码

感谢分享

感谢分享

感谢分享