文档反馈-MindSpore

你好，在`Model`类的初始化流程，会检查kwargs的范围是否在`['loss_scale_manager', 'keep_batchnorm_fp32', 'boost_config_dict', 'acc_level']`里：

```
# train/model.py
class Model:
    def __init__(..., **kwargs)
        ...
        self._check_kwargs(kwargs)
        ...

def _check_kwargs(self, kwargs):
        for arg in kwargs:
            if arg not in ['loss_scale_manager', 'keep_batchnorm_fp32', 'boost_config_dict', 'acc_level']:
                raise ValueError(f"The argument in 'kwargs' should be 'loss_scale_manager' or "
                                 f"'keep_batchnorm_fp32' or 'boost_config_dict' or 'acc_level', but got '{arg}'.")
```

在`build_train_network`函数里，会检查kwargs的范围是否在`['cast_model_type', 'keep_batchnorm_fp32', 'loss_scale_manager']`里
```
# train/amp.py
def _check_kwargs(key_words):
    """Check kwargs."""
    for arg in key_words:
        if arg not in ['cast_model_type', 'keep_batchnorm_fp32', 'loss_scale_manager']:
            raise ValueError(f"Unsupported arg '{arg}'")
    ...

def build_train_network(..., **kwargs):
    ...
    _check_kwargs(kwargs)
    ...
```

1、你说的`_check_for_graph_cell`函数，是不是指的是`build_train_network`里调用的`_check_kwargs`函数？
2、`build_train_network`里不会调用`Model`的`_check_kwargs`方法
3、`Model`虽然会调用`build_train_network`，但是调用`build_train_network`传的kwargs能保证符合`build_train_network`的要求

MindSpore/docs

内容风险标识

评论 (5)

MindSpore/docs .gitee-modal { width: 500px !important; }

内容风险标识