我正在实现将一个pytorch框架下的项目（已验证可以正常运行）转成mindspore框架下，在进行网络训练时，报了以下错误，请问这个应该怎么解决呢？有没有大佬解决过类似的问题？ 
Hardware Environment(Ascend/GPU/CPU) / 硬件环境: GPU cuda11.6
Software Environment / 软件环境 (Mandatory / 必填):
- MindSpore version (e.g., 1.7.0.Bxxx) : 2.2.10
- Python version (e.g., Python 3.7.5) : 3.9.18
- OS platform and distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 20.04 x86_64 GNU/Linux
- GCC/Compiler version (if compiled from source): 8.5.0
预期结果是输出:
而实际上只输出了前两行就没了
mindspore框架下的代码如下：
以下是pytorch版的源码
点赞
收藏
回复
分享
举报
已解决
我正在实现将一个pytorch框架下的项目（已验证可以正常运行）转成mindspore框架下，在进行网络训练时，报了以下错误，请问这个应该怎么解决呢？有没有大佬解决过类似的问题？
Hardware Environment(Ascend/GPU/CPU) / 硬件环境: GPU cuda11.6
Software Environment / 软件环境 (Mandatory / 必填):
预期结果是输出:
而实际上只输出了前两行就没了
mindspore框架下的代码如下：
def _train(self, loader_tr, clf_group: dict, clf_name='train', soft_target=False, log_show=True): """Represents one epoch. :param loader_tr: (:obj:`torch.utils.data.DataLoader`) The training data wrapped in DataLoader. Accuracy and loss in the each iter will be recorded. """ iter_out = self.args.out_iter_freq loss_list = [] right_count_list = [] samples_per_batch = [] # clf_group['clf'].train() clf_group['clf'].set_train(True) optimizer = clf_group['optimizer'] def train_step(data, label): print('train_step1') (loss, out), grads = grad_fn(data, label) print('train_step2') optimizer(grads) print('train_step3') return loss,out def forward_fn(data, label): print('forward_fn1') out, _, _ = self.clf(data) print('forward_fn2') loss = mindspore.ops.cross_entropy(out, label) print('forward_fn3') return loss, out grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True) for batch_idx, (x, y, _, _) in enumerate(loader_tr): squeeze = mindspore.ops.Squeeze(1) x = squeeze(x) if soft_target: y = y.astype(np.float32) else: y = y.astype(np.int32) print('00',type(x),type(y),x.shape,y.shape) loss,out = train_step(x, y) print(11) loss_list.append(loss.item()) if soft_target: pred_compare = out.max(axis=1, return_indices=True)[1] y_compare = y.max(axis=1, return_indices=True)[1] right_count_list.append((pred_compare == y_compare).sum().item()) samples_per_batch.append(len(y)) else: pred = out.max(axis=1, return_indices=True)[1] right_count_list.append((pred == y).sum().item()) samples_per_batch.append(len(y))以下是pytorch版的源码
def _train(self, loader_tr, clf_group: dict, clf_name='train', soft_target=False, log_show=True): """Represents one epoch. :param loader_tr: (:obj:`torch.utils.data.DataLoader`) The training data wrapped in DataLoader. Accuracy and loss in the each iter will be recorded. """ iter_out = self.args.out_iter_freq loss_list = [] right_count_list = [] samples_per_batch = [] clf_group['clf'].train() for batch_idx, (x, y, _, _) in enumerate(loader_tr): x, y = x.cuda(), y.cuda() clf_group['optimizer'].zero_grad() out, _, _ = self.clf(x) if soft_target: pred_compare = out.max(1)[1] y_compare = y.max(1)[1] right_count_list.append((pred_compare == y_compare).sum().item()) samples_per_batch.append(len(y)) else: pred = out.max(1)[1] right_count_list.append((pred == y).sum().item()) samples_per_batch.append(len(y)) loss = F.cross_entropy(out, y) loss_list.append(loss.item()) loss.backward() clf_group['optimizer'].step() clf_group['scheduler'].step()