238 lines
13 KiB
Markdown
238 lines
13 KiB
Markdown
# MMSegmentation 模型部署
|
||
|
||
- [MMSegmentation 模型部署](#mmsegmentation-模型部署)
|
||
- [安装](#安装)
|
||
- [安装 mmseg](#安装-mmseg)
|
||
- [安装 mmdeploy](#安装-mmdeploy)
|
||
- [模型转换](#模型转换)
|
||
- [模型规范](#模型规范)
|
||
- [模型推理](#模型推理)
|
||
- [后端模型推理](#后端模型推理)
|
||
- [SDK 模型推理](#sdk-模型推理)
|
||
- [模型支持列表](#模型支持列表)
|
||
- [注意事项](#注意事项)
|
||
|
||
______________________________________________________________________
|
||
|
||
[MMSegmentation](https://github.com/open-mmlab/mmsegmentation/tree/1.x) 又称`mmseg`，是一个基于 PyTorch 的开源对象分割工具箱。它是 [OpenMMLab](https://openmmlab.com/) 项目的一部分。
|
||
|
||
## 安装
|
||
|
||
### 安装 mmseg
|
||
|
||
请参考[官网安装指南](https://mmsegmentation.readthedocs.io/en/1.x/get_started.html)。
|
||
|
||
### 安装 mmdeploy
|
||
|
||
mmdeploy 有以下几种安装方式:
|
||
|
||
**方式一：** 安装预编译包
|
||
|
||
> 待 mmdeploy 正式发布 1.x，再补充
|
||
|
||
**方式二：** 一键式脚本安装
|
||
|
||
如果部署平台是 **Ubuntu 18.04 及以上版本**， 请参考[脚本安装说明](../01-how-to-build/build_from_script.md)，完成安装过程。
|
||
比如，以下命令可以安装 mmdeploy 以及配套的推理引擎——`ONNX Runtime`.
|
||
|
||
```shell
|
||
git clone --recursive -b 1.x https://github.com/open-mmlab/mmdeploy.git
|
||
cd mmdeploy
|
||
python3 tools/scripts/build_ubuntu_x64_ort.py $(nproc)
|
||
export PYTHONPATH=$(pwd)/build/lib:$PYTHONPATH
|
||
export LD_LIBRARY_PATH=$(pwd)/../mmdeploy-dep/onnxruntime-linux-x64-1.8.1/lib/:$LD_LIBRARY_PATH
|
||
```
|
||
|
||
**说明**:
|
||
|
||
- 把 `$(pwd)/build/lib` 添加到 `PYTHONPATH`，目的是为了加载 mmdeploy SDK python 包 `mmdeploy_python`，在章节 [SDK模型推理](#sdk模型推理)中讲述其用法。
|
||
- 在[使用 ONNX Runtime推理后端模型](#后端模型推理)时，需要加载自定义算子库，需要把 ONNX Runtime 库的路径加入环境变量 `LD_LIBRARY_PATH`中。
|
||
**方式三：** 源码安装
|
||
|
||
在方式一、二都满足不了的情况下，请参考[源码安装说明](../01-how-to-build/build_from_source.md) 安装 mmdeploy 以及所需推理引擎。
|
||
|
||
## 模型转换
|
||
|
||
你可以使用 [tools/deploy.py](https://github.com/open-mmlab/mmdeploy/tree/1.x/tools/deploy.py) 把 mmseg 模型一键式转换为推理后端模型。
|
||
该工具的详细使用说明请参考[这里](https://github.com/open-mmlab/mmdeploy/tree/1.x/docs/en/02-how-to-run/convert_model.md#usage).
|
||
|
||
以下，我们将演示如何把 `unet` 转换为 onnx 模型。
|
||
|
||
```shell
|
||
cd mmdeploy
|
||
|
||
# download unet model from mmseg model zoo
|
||
mim download mmsegmentation --config unet-s5-d16_fcn_4xb4-160k_cityscapes-512x1024 --dest .
|
||
|
||
# convert mmseg model to onnxruntime model with dynamic shape
|
||
python tools/deploy.py \
|
||
configs/mmseg/segmentation_onnxruntime_dynamic.py \
|
||
unet-s5-d16_fcn_4xb4-160k_cityscapes-512x1024.py \
|
||
fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes_20211210_145204-6860854e.pth \
|
||
demo/resources/cityscapes.png \
|
||
--work-dir mmdeploy_models/mmseg/ort \
|
||
--device cpu \
|
||
--show \
|
||
--dump-info
|
||
```
|
||
|
||
转换的关键之一是使用正确的配置文件。项目中已内置了各后端部署[配置文件](https://github.com/open-mmlab/mmdeploy/tree/1.x/configs/mmseg)。
|
||
文件的命名模式是：
|
||
|
||
```
|
||
segmentation_{backend}-{precision}_{static | dynamic}_{shape}.py
|
||
```
|
||
|
||
其中：
|
||
|
||
- **{backend}:** 推理后端名称。比如，onnxruntime、tensorrt、pplnn、ncnn、openvino、coreml 等等
|
||
- **{precision}:** 推理精度。比如，fp16、int8。不填表示 fp32
|
||
- **{static | dynamic}:** 动态、静态 shape
|
||
- **{shape}:** 模型输入的 shape 或者 shape 范围
|
||
|
||
在上例中，你也可以把 `unet` 转为其他后端模型。比如使用`segmentation_tensorrt-fp16_dynamic-512x1024-2048x2048.py`，把模型转为 tensorrt-fp16 模型。
|
||
|
||
```{tip}
|
||
当转 tensorrt 模型时, --device 需要被设置为 "cuda"
|
||
```
|
||
|
||
## 模型规范
|
||
|
||
在使用转换后的模型进行推理之前，有必要了解转换结果的结构。 它存放在 `--work-dir` 指定的路路径下。
|
||
|
||
上例中的`mmdeploy_models/mmseg/ort`，结构如下：
|
||
|
||
```
|
||
mmdeploy_models/mmseg/ort
|
||
├── deploy.json
|
||
├── detail.json
|
||
├── end2end.onnx
|
||
└── pipeline.json
|
||
```
|
||
|
||
重要的是：
|
||
|
||
- **end2end.onnx**: 推理引擎文件。可用 ONNX Runtime 推理
|
||
- \***.json**: mmdeploy SDK 推理所需的 meta 信息
|
||
|
||
整个文件夹被定义为**mmdeploy SDK model**。换言之，**mmdeploy SDK model**既包括推理引擎，也包括推理 meta 信息。
|
||
|
||
## 模型推理
|
||
|
||
### 后端模型推理
|
||
|
||
以上述模型转换后的 `end2end.onnx` 为例，你可以使用如下代码进行推理：
|
||
|
||
```python
|
||
from mmdeploy.apis.utils import build_task_processor
|
||
from mmdeploy.utils import get_input_shape, load_config
|
||
import torch
|
||
|
||
deploy_cfg = 'configs/mmseg/segmentation_onnxruntime_dynamic.py'
|
||
model_cfg = './unet-s5-d16_fcn_4xb4-160k_cityscapes-512x1024.py'
|
||
device = 'cpu'
|
||
backend_model = ['./mmdeploy_models/mmseg/ort/end2end.onnx']
|
||
image = './demo/resources/cityscapes.png'
|
||
|
||
# read deploy_cfg and model_cfg
|
||
deploy_cfg, model_cfg = load_config(deploy_cfg, model_cfg)
|
||
|
||
# build task and backend model
|
||
task_processor = build_task_processor(model_cfg, deploy_cfg, device)
|
||
model = task_processor.build_backend_model(backend_model)
|
||
|
||
# process input image
|
||
input_shape = get_input_shape(deploy_cfg)
|
||
model_inputs, _ = task_processor.create_input(image, input_shape)
|
||
|
||
# do model inference
|
||
with torch.no_grad():
|
||
result = model.test_step(model_inputs)
|
||
|
||
# visualize results
|
||
task_processor.visualize(
|
||
image=image,
|
||
model=model,
|
||
result=result[0],
|
||
window_name='visualize',
|
||
output_file='./output_segmentation.png')
|
||
```
|
||
|
||
### SDK 模型推理
|
||
|
||
你也可以参考如下代码，对 SDK model 进行推理：
|
||
|
||
```python
|
||
from mmdeploy_python import Segmentor
|
||
import cv2
|
||
import numpy as np
|
||
|
||
img = cv2.imread('./demo/resources/cityscapes.png')
|
||
|
||
# create a classifier
|
||
segmentor = Segmentor(model_path='./mmdeploy_models/mmseg/ort', device_name='cpu', device_id=0)
|
||
# perform inference
|
||
seg = segmentor(img)
|
||
|
||
# visualize inference result
|
||
## random a palette with size 256x3
|
||
palette = np.random.randint(0, 256, size=(256, 3))
|
||
color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8)
|
||
for label, color in enumerate(palette):
|
||
color_seg[seg == label, :] = color
|
||
# convert to BGR
|
||
color_seg = color_seg[..., ::-1]
|
||
img = img * 0.5 + color_seg * 0.5
|
||
img = img.astype(np.uint8)
|
||
cv2.imwrite('output_segmentation.png', img)
|
||
```
|
||
|
||
除了python API，mmdeploy SDK 还提供了诸如 C、C++、C#、Java等多语言接口。
|
||
你可以参考[样例](https://github.com/open-mmlab/mmdeploy/tree/1.x/demo)学习其他语言接口的使用方法。
|
||
|
||
## 模型支持列表
|
||
|
||
| Model | OnnxRuntime | TensorRT | ncnn | PPLNN | OpenVino |
|
||
| :------------------------------------------------------------------------------------------------------- | :---------: | :------: | :--: | :---: | :------: |
|
||
| [FCN](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/fcn) | Y | Y | Y | Y | Y |
|
||
| [PSPNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/pspnet)[\*](#static_shape) | Y | Y | Y | Y | Y |
|
||
| [DeepLabV3](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/deeplabv3) | Y | Y | Y | Y | Y |
|
||
| [DeepLabV3+](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/deeplabv3plus) | Y | Y | Y | Y | Y |
|
||
| [Fast-SCNN](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/fastscnn)[\*](#static_shape) | Y | Y | N | Y | Y |
|
||
| [UNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/unet) | Y | Y | Y | Y | Y |
|
||
| [ANN](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/ann)[\*](#static_shape) | Y | Y | N | N | N |
|
||
| [APCNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/apcnet) | Y | Y | Y | N | N |
|
||
| [BiSeNetV1](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/bisenetv1) | Y | Y | Y | N | Y |
|
||
| [BiSeNetV2](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/bisenetv2) | Y | Y | Y | N | Y |
|
||
| [CGNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/cgnet) | Y | Y | Y | N | Y |
|
||
| [DMNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/dmnet) | Y | N | N | N | N |
|
||
| [DNLNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/dnlnet) | Y | Y | Y | N | Y |
|
||
| [EMANet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/emanet) | Y | Y | N | N | Y |
|
||
| [EncNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/encnet) | Y | Y | N | N | Y |
|
||
| [ERFNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/erfnet) | Y | Y | Y | N | Y |
|
||
| [FastFCN](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/fastfcn) | Y | Y | Y | N | Y |
|
||
| [GCNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/gcnet) | Y | Y | N | N | N |
|
||
| [ICNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/icnet)[\*](#static_shape) | Y | Y | N | N | Y |
|
||
| [ISANet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/isanet)[\*](#static_shape) | Y | Y | N | N | Y |
|
||
| [NonLocal Net](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/nonlocal_net) | Y | Y | Y | N | Y |
|
||
| [OCRNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/ocrnet) | Y | Y | Y | N | Y |
|
||
| [PointRend](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/point_rend)[\*](#static_shape) | Y | Y | N | N | N |
|
||
| [Semantic FPN](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/sem_fpn) | Y | Y | Y | N | Y |
|
||
| [STDC](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/stdc) | Y | Y | Y | N | Y |
|
||
| [UPerNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/upernet)[\*](#static_shape) | Y | Y | N | N | N |
|
||
| [DANet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/danet) | Y | Y | N | N | Y |
|
||
| [Segmenter](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/segmenter)[\*](#static_shape) | Y | Y | Y | N | Y |
|
||
| [SegFormer](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/segformer)[\*](#static_shape) | Y | Y | N | N | Y |
|
||
| [SETR](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/setr) | Y | N | N | N | Y |
|
||
| [CCNet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/ccnet) | N | N | N | N | N |
|
||
| [PSANet](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/psanet) | N | N | N | N | N |
|
||
| [DPT](https://github.com/open-mmlab/mmsegmentation/tree/1.x/configs/dpt) | N | N | N | N | N |
|
||
|
||
## 注意事项
|
||
|
||
- 所有 mmseg 模型仅支持 "whole" 推理模式。
|
||
|
||
- <i id=“static_shape”>PSPNet，Fast-SCNN</i> 仅支持静态输入，因为多数推理框架的 [nn.AdaptiveAvgPool2d](https://github.com/open-mmlab/mmsegmentation/blob/0c87f7a0c9099844eff8e90fa3db5b0d0ca02fee/mmseg/models/decode_heads/psp_head.py#L38) 不支持动态输入。
|
||
|
||
- 对于仅支持静态形状的模型，应使用静态形状的部署配置文件，例如 `configs/mmseg/segmentation_tensorrt_static-1024x2048.py`
|