[Feature]Add paddleslim ACT. #3457
Conversation
deploy/slim/act/readme_ch.md
Outdated
| #### 3.1 准备环境 | ||
|
|
||
| - PaddlePaddle >= 2.4 （可从[Paddle官网](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/install/pip/linux-pip.html)下载安装） | ||
| - PaddleSlim >= 2.4 |
There was a problem hiding this comment.
paddle和slim都是2.5，套件使用develop，pr合入develop，后cherry-pick到最新分支
deploy/slim/act/run_seg.py
Outdated
| config['params_filename']) # get the name of forward input | ||
| train_dataloader = reader_wrapper(train_loader, input_name) | ||
|
|
||
| rank_id = paddle.distributed.get_rank() |
deploy/slim/act/readme.md
Outdated
| | 模型 | 策略 | Total IoU (%) | CPU耗时(ms)<br>thread=10<br>mkldnn=on| Nvidia GPU耗时(ms)<br>TRT=off| 配置文件 | Inference模型 | | ||
| |:-----:|:-----:|:----------:|:---------:| :------:|:------:|:------:| | ||
| | OCRNet_HRNetW48 |Baseline |todo| todo| todo|todo|todo| | ||
| | OCRNet_HRNetW48 | 量化训练QAT |todo| todo| todo|todo|todo| |
There was a problem hiding this comment.
只有量化训练QAT 还是 量化训练QAT结合蒸馏？
deploy/slim/act/readme.md
Outdated
|
|
||
| ## 2.Benchmark | ||
|
|
||
| | 模型 | 策略 | Total IoU (%) | CPU耗时(ms)<br>thread=10<br>mkldnn=on| Nvidia GPU耗时(ms)<br>TRT=off| 配置文件 | Inference模型 | |
There was a problem hiding this comment.
这个表格需要展示压缩前后的精度和推理速度，现在表头的字段不对啊
There was a problem hiding this comment.
对于不同配置分成了两行来对比，一行是baseline，一行QAT，后续增加不同的配置可以增加多行，方便不同配置之间的对比。
deploy/slim/act/readme.md
Outdated
|
|
||
| - 示例数据集: cityscapes数据集的一个子集，用于快速跑通压缩和推理流程，不能用该数据集复现 benchmark 表中的压缩效果。[下载链接](https://bj.bcebos.com/v1/paddle-slim-models/data/mini_cityscapes/mini_cityscapes.tar) | ||
|
|
||
| 准备好数据后，需要放入到`deploy/slim/act`目录下。 |
There was a problem hiding this comment.
怎么放？ 整个文档需要人家可以执行复现，比如文档甩给qa可以帮你测试完成，可以再整体优化下
There was a problem hiding this comment.
已经修改内容，并已经向QA提测。
| @@ -0,0 +1,319 @@ | |||
| # Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. | |||
| return miou | ||
|
|
||
|
|
||
| def reader_wrapper(reader, input_name): |
There was a problem hiding this comment.
开发者不太清楚这个wrapper的作用，需要说明act对dataset的要求，这部分还挺关键
There was a problem hiding this comment.
act对dataset的要求和套件一致，因为数据的配置使用的seg的配置文件，通过segbuilder加载之后，使用dataloader获取数据。
|
|
||
| ### 3. 量化蒸馏训练精度很低？ | ||
|  | ||
| **A**：去除量化训练的输出结果，重新运行一次，这是由于网络训练到局部极值点导致。 |
| @@ -0,0 +1 @@ | |||
| 请参考[语义分割模型自动压缩示例](../../../../deploy/slim/act/readme.md)了解更多基于PaddleSeg接入ACT的方式。 | |||
There was a problem hiding this comment.
在seg的首页readme里面，可以加下Auto Compression，链接到这里面来
PR types
New features
PR changes
Others
Description
Add paddleslim ACT function on ppliteseg as an example.