mirror of https://github.com/alibaba/MNN.git
91 lines
1.3 KiB
Markdown
91 lines
1.3 KiB
Markdown
|
## optim.Optimizer
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
class Optimizer
|
|||
|
|
```
|
|||
|
|
Optimizer是一个优化器基类`Optimizer`，`SGD`和`ADAM`都是该类的具体实现
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `Optimizer()`
|
|||
|
|
创建一个空Optimizer
|
|||
|
|
|
|||
|
|
*在实际使用中创建空Optimizer没有意义，请使用`optim.SGD`或`optim.ADAM`来创建优化器实例*
|
|||
|
|
|
|||
|
|
参数：
|
|||
|
|
- `None`
|
|||
|
|
|
|||
|
|
返回：Optimizer对象
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
返回类型：`Optimizer`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `learning_rate`
|
|||
|
|
|
|||
|
|
获取和设置优化器的学习率
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`float`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `momentum`
|
|||
|
|
|
|||
|
|
获取和设置优化器的动量
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`float`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `momentum2`
|
|||
|
|
|
|||
|
|
获取并设置优化器的第二个动量，*只有ADAM优化器才有该属性*
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`float`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `weight_decay`
|
|||
|
|
|
|||
|
|
获取并设置优化器的权重衰减因子
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`float`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `eps`
|
|||
|
|
|
|||
|
|
获取并设置优化器的eps系数，*只有ADAM优化器才有该属性*
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`float`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `regularization_method`
|
|||
|
|
|
|||
|
|
获取并设置优化器的正则化方法
|
|||
|
|
|
|||
|
|
属性类型：读写
|
|||
|
|
|
|||
|
|
类型：`RegularizationMethod`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `step(loss)`
|
|||
|
|
反向传播以获得参数的梯度，并使用其相应的梯度更新参数
|
|||
|
|
|
|||
|
|
参数：
|
|||
|
|
- `loss:Var` 当前的损失函数值
|
|||
|
|
|
|||
|
|
返回：是否更新成功
|
|||
|
|
|
|||
|
|
返回类型：`bool`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
示例
|
|||
|
|
|
|||
|
|
请参考[`optim`](optim.md)
|