mirror of https://github.com/alibaba/MNN.git
44 lines
1.1 KiB
Markdown
44 lines
1.1 KiB
Markdown
|
## compress
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
module compress
|
|||
|
|
```
|
|||
|
|
compress模块用来做`Quantization-Aware-Training(QAT)`训练量化，提供了训练量化的接口
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `compress.Feature_Scale_Method`
|
|||
|
|
对特征的量化方式，可以针对整个特征进行量化，也可以针对每个channel进行量化
|
|||
|
|
- 类型：`Enum`
|
|||
|
|
- 枚举值：
|
|||
|
|
- `PER_TENSOR`
|
|||
|
|
- `PER_CHANNEL`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `compress.Scale_Update_Method`
|
|||
|
|
scale的更新方式
|
|||
|
|
- 类型：`Enum`
|
|||
|
|
- 枚举值：
|
|||
|
|
- `MAXIMUM`
|
|||
|
|
- `MOVING_AVERAGE`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### `train_quant(module, |quant_bits, feature_scale_method, scale_update_method)`
|
|||
|
|
|
|||
|
|
训练量化
|
|||
|
|
|
|||
|
|
参数：
|
|||
|
|
- `module` 待训练模型
|
|||
|
|
- `quant_bits` 量化位数，默认为 `8`
|
|||
|
|
- `feature_scale_method` 特征的量化方式，默认为 `PER_TENSOR`
|
|||
|
|
- `scale_update_method` scale的更新方式，默认为 `MOVING_AVERAGE`
|
|||
|
|
|
|||
|
|
返回：是否成功
|
|||
|
|
|
|||
|
|
返回类型：`bool`
|
|||
|
|
|
|||
|
|
示例
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
# args are self-explained
|
|||
|
|
nn.compress.train_quant(module, quant_bits = 8, feature_scale_method = Feature_Scale_Method.PER_TENSOR, scale_update_method = Scale_Update_Method.MOVING_AVERAGE)
|
|||
|
|
```
|