add new model

0d97cc8c · Sugon_ldc · 0d97cc8c · 0d97cc8c · 0d97cc8c · 0d97cc8c
Commit 0d97cc8c authored Jun 07, 2023 by Sugon_ldc
20 changed files
--- a/configs/segformer/segformer_b5_cityscapes_1024x1024_160k.yml
+++ b/configs/segformer/segformer_b5_cityscapes_1024x1024_160k.yml
+_base_: '../_base_/cityscapes_1024x1024.yml'
+
+batch_size: 1
+iters: 160000
+
+model:
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B5
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b5.tar.gz
+  embedding_dim: 768
+  num_classes: 19
+
+optimizer:
+  _inherited_: False
+  type: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  weight_decay: 0.01
+
+lr_scheduler:
+  type: PolynomialDecay
+  learning_rate: 0.00006
+  power: 1
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+  coef: [1]
+
+test_config:
+    is_slide: True
+    crop_size: [1024, 1024]
+    stride: [768, 768]
--- a/configs/segformer/segformer_b5_cityscapes_1024x512_160k.yml
+++ b/configs/segformer/segformer_b5_cityscapes_1024x512_160k.yml
+_base_: '../_base_/cityscapes.yml'
+
+batch_size: 1
+iters: 160000
+
+model:
+  type: SegFormer
+  backbone:
+    type: MixVisionTransformer_B5
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b5.tar.gz
+  embedding_dim: 768
+  num_classes: 19
+
+optimizer:
+  _inherited_: False
+  type: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  weight_decay: 0.01
+
+lr_scheduler:
+  type: PolynomialDecay
+  learning_rate: 0.00006
+  power: 1
+
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+  coef: [1]
--- a/configs/segmenter/README.md
+++ b/configs/segmenter/README.md
+# Segmenter: Transformer for Semantic Segmentation
+
+## Reference
+
+> Strudel, Robin, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. "Segmenter: Transformer for Semantic Segmentation." In Proceedings of the IEEE International Conference on Computer Vision, pp. 7262-7272. 2021.
+
+## Performance
+
+### ADE20k
+
+| Model | Backbone | Head | Patch Size | Resolution | Training Iters | mIoU (slice) | mIoU (flip)  | Links |
+| :-:   | :-:      | :-:  | :-:        | :-:        | :-:            | :-:          | :-:          | :-:  |
+| Segmenter | ViT small | Linear  | 16 | 512*512 | 160000 | 45.48 | 45.69 | [model](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_small_linear_ade20k_512x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_small_linear_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=4dc954a9b774e4807c07c511c04ce0f6) |
+| Segmenter | ViT small | Mask    | 16 | 512*512 | 160000 | 45.15 | 45.41 |  [model](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_small_mask_ade20k_512x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_small_mask_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=0fdd5191ecec56bbdf08259cc6c32a21) |
+| Segmenter | ViT base  | Linear  | 16 | 512*512 | 160000 | 48.13 | 48.31 |  [model](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_base_linear_ade20k_512x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_base_linear_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/index?id=992f38b3f937de87dc74a888d217f53e) |
+| Segmenter | ViT base  | Mask    | 16 | 512*512 | 160000 | 48.49 | 48.61 |  [model](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_base_mask_ade20k_512x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ade20k/segmenter_vit_base_mask_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=16a7380069b6435bdf6e566dcc7f4a6b) |
--- a/configs/segmenter/segmenter_vit_base_linear_ade20k_512x512_160k.yml
+++ b/configs/segmenter/segmenter_vit_base_linear_ade20k_512x512_160k.yml
+_base_: '../_base_/ade20k.yml'
+
+batch_size: 2
+iters: 160000
+
+model:
+  type: LinearSegmenter
+  backbone:
+    type: VisionTransformer
+    img_size: 512
+    patch_size: 16
+    embed_dim: 768
+    depth: 12
+    num_heads: 12
+    mlp_ratio: 4
+    qkv_bias: True
+    drop_rate: 0.0
+    drop_path_rate: 0.1
+    final_norm: True
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/pretrained_models/vit_base_patch16_384_augreg.tar.gz
+
+val_dataset:
+  transforms:
+    - type: ResizeByShort
+      short_size: 512
+    - type: Normalize
+
+optimizer:
+  weight_decay: 0.0
+
+lr_scheduler:
+  learning_rate: 0.001
+  end_lr: 1.0e-05
+
+test_config:
+  is_slide: True
+  crop_size: [512, 512]
+  stride: [512, 512]
--- a/configs/segmenter/segmenter_vit_base_mask_ade20k_512x512_160k.yml
+++ b/configs/segmenter/segmenter_vit_base_mask_ade20k_512x512_160k.yml
+_base_: './segmenter_vit_base_linear_ade20k_512x512_160k.yml'
+
+model:
+  type: MaskSegmenter
+  h_embed_dim: 768
+  h_depth: 2
+  h_num_heads: 12
+  h_mlp_ratio: 4
+  h_drop_rate: 0.0
+  h_drop_path_rate: 0.1
--- a/configs/segmenter/segmenter_vit_small_linear_ade20k_512x512_160k.yml
+++ b/configs/segmenter/segmenter_vit_small_linear_ade20k_512x512_160k.yml
+_base_: './segmenter_vit_base_linear_ade20k_512x512_160k.yml'
+
+model:
+  type: LinearSegmenter
+  backbone:
+    type: VisionTransformer
+    img_size: 512
+    patch_size: 16
+    embed_dim: 384
+    depth: 12
+    num_heads: 6
+    mlp_ratio: 4
+    qkv_bias: True
+    drop_rate: 0.0
+    drop_path_rate: 0.1
+    final_norm: True
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/pretrained_models/vit_small_patch16_384_augreg.tar.gz
--- a/configs/segmenter/segmenter_vit_small_mask_ade20k_512x512_160k.yml
+++ b/configs/segmenter/segmenter_vit_small_mask_ade20k_512x512_160k.yml
+_base_: './segmenter_vit_small_linear_ade20k_512x512_160k.yml'
+
+model:
+  type: MaskSegmenter
+  h_embed_dim: 384
+  h_depth: 2
+  h_num_heads: 6
+  h_mlp_ratio: 4
+  h_drop_rate: 0.0
+  h_drop_path_rate: 0.1
--- a/configs/segnet/README.md
+++ b/configs/segnet/README.md
+# SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
+
+## Reference
+> Badrinarayanan, Vijay, Alex Kendall, and Roberto Cipolla. "Segnet: A deep convolutional encoder-decoder architecture for image segmentation." IEEE transactions on pattern analysis and machine intelligence 39, no. 12 (2017): 2481-2495.
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (flip) | mIoU (ms+flip) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|SegNet|-|1024x512|80000|60.09%|60.88%|61.85%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segnet_cityscapes_1024x512_80k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segnet_cityscapes_1024×512_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=cb3abc86f6a3ebcd2d3033a68b23162d)
--- a/configs/segnet/segnet_cityscapes_1024x512_80k.yml
+++ b/configs/segnet/segnet_cityscapes_1024x512_80k.yml
+_base_: '../_base_/cityscapes.yml'
+
+batch_size: 2
+iters: 80000
+
+model:
+  type: SegNet
+  pretrained: Null
--- a/configs/setr/README.md
+++ b/configs/setr/README.md
+# Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers
+
+## Reference
+
+> Zheng, Sixiao, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu et al. "Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers." arXiv preprint arXiv:2012.15840 (2020).
+
+## Performance
+
+### CityScapes
+
+| Model | Backbone | Head | Resolution | Training Iters | mIoU(slice) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|Segmentation Transformer|Vision Transformer|Naive|769x769|40000|77.29%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_naive_large_cityscapes_769x769_40k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_naive_large_cityscapes_769x769_40k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=e21e3b4721366602a9a63c551108da1c)|
+|Segmentation Transformer|Vision Transformer|PUP|769x769|40000|78.08%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_pup_large_cityscapes_769x769_40k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_pup_large_cityscapes_769x769_40k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=c25fdd4ac6221704d278b09f19ddf970) |
+|Segmentation Transformer|Vision Transformer|MLA|769x769|40000|76.52%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_mla_large_cityscapes_769x769_40k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/setr_mla_large_cityscapes_769x769_40k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=993754909236b762b5276897ebec9c6d) |
--- a/configs/setr/setr_mla_large_cityscapes_769x769_40k.yml
+++ b/configs/setr/setr_mla_large_cityscapes_769x769_40k.yml
+_base_: '../_base_/cityscapes_769x769_setr.yml'
+
+model:
+  type: SegmentationTransformer
+  backbone:
+    type: ViT_large_patch16_384
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/vit_large_patch16_384.tar.gz
+  num_classes: 19
+  backbone_indices: [5, 11, 17, 23]
+  head: mla
+  align_corners: True
+
+optimizer:
+  weight_decay: 0.0
+
+lr_scheduler:
+  end_lr: 1.0e-4
+
+iters: 40000
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+  coef: [1, 0.4, 0.4, 0.4, 0.4]
+
+test_config:
+    is_slide: True
+    crop_size: [769, 769]
+    stride: [512, 512]
--- a/configs/setr/setr_naive_large_cityscapes_769x769_40k.yml
+++ b/configs/setr/setr_naive_large_cityscapes_769x769_40k.yml
+_base_: '../_base_/cityscapes_769x769_setr.yml'
+
+model:
+  type: SegmentationTransformer
+  backbone:
+    type: ViT_large_patch16_384
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/vit_large_patch16_384.tar.gz
+  num_classes: 19
+  backbone_indices: [9, 14, 19, 23]
+  head: naive
+  align_corners: True
+
+optimizer:
+  weight_decay: 0.0
+
+lr_scheduler:
+  end_lr: 1.0e-4
+
+iters: 40000
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+  coef: [1, 0.4, 0.4, 0.4]
+
+test_config:
+    is_slide: True
+    crop_size: [769, 769]
+    stride: [512, 512]
--- a/configs/setr/setr_pup_large_cityscapes_769x769_40k.yml
+++ b/configs/setr/setr_pup_large_cityscapes_769x769_40k.yml
+_base_: '../_base_/cityscapes_769x769_setr.yml'
+
+model:
+  type: SegmentationTransformer
+  backbone:
+    type: ViT_large_patch16_384
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/vit_large_patch16_384.tar.gz
+  num_classes: 19
+  backbone_indices: [9, 14, 19, 23]
+  head: pup
+  align_corners: True
+
+optimizer:
+  weight_decay: 0.0
+
+lr_scheduler:
+  end_lr: 1.0e-4
+
+iters: 40000
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+  coef: [1, 0.4, 0.4, 0.4, 0.4]
+
+test_config:
+    is_slide: True
+    crop_size: [769, 769]
+    stride: [512, 512]
--- a/configs/sfnet/README.md
+++ b/configs/sfnet/README.md
+# Semantic Flow for Fast and Accurate Scene Parsing
+
+## Reference
+
+> Xiangtai Li, Ansheng You, Zhen Zhu, Houlong Zhao, Maoke Yang, Kuiyuan Yang, Shaohua Tan, Yunhai Tong:
+Semantic Flow for Fast and Accurate Scene Parsing. ECCV (1) 2020: 775-793 .
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (flip) | mIoU (ms+flip) | Links |
+|-|-|-|-|-|-|-|-|
+|SFNet|ResNet18_OS8|1024x1024|80000|78.72%|79.11%|79.28%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/sfnet_resnet18_os8_cityscapes_1024x1024_80k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/sfnet_resnet18_os8_cityscapes_1024x1024_80k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=0d790ad96282048b136342fcebb08d14)|
+|SFNet|ResNet50_OS8|1024x1024|80000|81.49%|81.63%|81.85%|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/sfnet_resnet50_os8_cityscapes_1024x1024_80k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/sfnet_resnet50_os8_cityscapes_1024x1024_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=d458349ec63ea8ccd6fae84afa8ea981)|
--- a/configs/sfnet/sfnet_resnet18_os8_cityscapes_1024x1024_80k.yml
+++ b/configs/sfnet/sfnet_resnet18_os8_cityscapes_1024x1024_80k.yml
+_base_: '../_base_/cityscapes_1024x1024.yml'
+
+batch_size: 4
+iters: 80000
+
+model:
+  type: SFNet
+  backbone:
+    type: ResNet18_vd
+    output_stride: 8
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet18_vd_ssld_v2.tar.gz
+  backbone_indices: [0, 1, 2, 3]
+
+train_dataset:
+  transforms:
+    - type: ResizeStepScaling
+      min_scale_factor: 0.5
+      max_scale_factor: 2.0
+      scale_step_size: 0.25
+    - type: RandomPaddingCrop
+      crop_size: [1024, 1024]
+    - type: RandomHorizontalFlip
+    - type: RandomDistort
+      brightness_range: 0.4
+      contrast_range: 0.4
+      saturation_range: 0.4
+    - type: Normalize
+
+loss:
+  types:
+    - type: OhemCrossEntropyLoss
+  coef: [1]
+
+optimizer:
+  type: sgd
+  momentum: 0.9
+  weight_decay: 0.0005
--- a/configs/sfnet/sfnet_resnet50_os8_cityscapes_1024x1024_80k.yml
+++ b/configs/sfnet/sfnet_resnet50_os8_cityscapes_1024x1024_80k.yml
+_base_: 'sfnet_resnet18_os8_cityscapes_1024x1024_80k.yml'
+
+batch_size: 2
+iters: 80000
+
+model:
+  type: SFNet
+  backbone:
+    type: ResNet50_vd
+    output_stride: 8
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet50_vd_ssld_v2.tar.gz
+  backbone_indices: [0, 1, 2, 3]
--- a/configs/smrt/README.md
+++ b/configs/smrt/README.md
+> 本目录下的配置文件用于飞桨模型选型工具（PaddleSMRT）。
+# PaddleSMRT
+
+## 一、项目介绍
+
+[PaddleSMRT](https://www.paddlepaddle.org.cn/smrt) 是飞桨结合产业落地经验推出的产业模型选型工具，在项目落地过程中，用户根据自身的实际情况，输入自己的需求，即可以得到对应在算法模型、部署硬件以及教程文档的信息。
+同时为了更加精准的推荐，增加了数据分析功能，用户上传自己的标注文件，系统可以自动分析数据特点，例如数据分布不均衡、小目标、密集型等，从而提供更加精准的模型以及优化策略，更好的符合场景的需求。
+
+
+本文档主要介绍PaddleSMRT在分割方向上是如何进行模型选型推荐，以及推荐模型的使用方法。
+
+## 二、数据介绍
+
+PaddleSMRT结合产业真实场景，通过比较算法效果，向用户推荐最适合的模型。目前PaddleSMRT覆盖工业质检、城市安防两大场景，下面介绍PaddleSMRT进行算法对比所使用的数据集。
+
+### 1. 新能源电池质检数据集
+
+数据集为新能源电池电池组件质检数据集，包含15021张图片，覆盖45种缺陷类型，例如掉胶，裂纹，划痕等。
+
+新能源电池数据展示图:
+
+<div align="center">
+  <img src="https://user-images.githubusercontent.com/48433081/169200335-c77d58e4-8916-46e4-be4b-eb7fe48a2f80.png"  width = "600" />  
+</div>
+
+数据集特点为：
+
+1. 类别分布均衡
+2. 属于小目标数据
+3. 非密集型数据
+
+### 2. 铝件质检数据集
+
+数据集为铝件生产过程中的质检数据集，包含11293张图片，覆盖5种缺陷类型，例如划伤，压伤，起皮等。
+
+铝件质检数据展示图:
+
+<div align="center">
+  <img src="https://user-images.githubusercontent.com/48433081/169200252-95a69964-0ae1-40bb-b2b9-2ba17f9bef64.png"  width = "600" />  
+</div>
+
+
+数据集特点为：
+
+1. 类别分布不均衡
+2. 属于小目标数据
+3. 非密集型数据
+
+
+## 三、推荐模型使用全流程
+
+通过飞桨官网的[模型选型工具](https://www.paddlepaddle.org.cn/smrt)明确需要使用的模型后，大家需要进行数据准备、模型训练、模型导出、模型部署，下面我们以一个例子进行简要说明。
+
+### 3.1 准备环境
+
+参考PaddleSeg[安装文档](../../docs/install_cn.md)安装PaddlePaddle、下载PaddleSeg代码、安装PaddleSeg依赖库。
+
+### 3.2 准备数据
+
+详细的数据准备方法，请参考[数据标注文档](../../docs/data/marker/marker_cn.md)和[数据配置文档](../../docs/data/custom/data_prepare_cn.md)。
+
+此处，我们准备了一个缺陷分割的数据集，点击[链接](https://paddle-smrt.bj.bcebos.com/data/seg/defect_data.zip)下载，或者执行如下命令下载。
+```
+wget https://paddle-smrt.bj.bcebos.com/data/seg/defect_data.zip
+```
+
+将下载的数据集解压，并存放到`PaddleSeg/data/defect_data`目录，如下。
+```
+PaddleSeg/data/defect_data
+├── Annotations
+├── JPEGImages
+├── test.txt
+├── train.txt
+└── val.txt
+```
+
+该数据集的原始图片保存在JPEGImages文件，标注图片保存在Annotations。
+
+train.txt、val.txt和test.txt分别标识用于训练、验证和测试的数据，其中内容如下，每一行的前面表示原始图像的相对路径、后面表示标注图像的相对路径。
+```
+JPEGImages/liewen_26.png Annotations/liewen_26.png
+JPEGImages/diaojiao_394.png Annotations/diaojiao_394.png
+```
+
+标注图像包含背景和3类缺陷目标，总共4类，分别标注的数值使0，1，2，3。
+
+### 3.3 准备配置文件
+
+PaddleSeg推荐使用配置文件的方式来训练、导出模型，简单方便。
+
+针对工业质检任务，我们为模型选型工具推荐的6个模型准备好了配置文件，存放在`PaddleSeg/configs/smrt`目录下。
+
+```
+PaddleSeg/configs/smrt
+├── base_cfg.yml  
+├── pp_liteseg_stdc1.yml  
+├── pp_liteseg_stdc2.yml  
+├── deeplabv3p_resnet50_os8.yml  
+├── ocrnet_hrnetw18.yml
+├── bisenetv2.yml
+└── sfnet_resnet18_os8.yml
+```
+
+其中，`base_cfg.yml`是公共配置，包括数据集、优化器、学习率，其他文件包含`base_cfg.yml`，额外定义了模型、损失函数。
+
+在其他应用中，大家可以根据实际情况修改上述配置文件中的字段，而且需要根据模型数据量调整配置文件中的超参，比如训练轮数iters、batch_size、学习率等。
+
+### 3.4 执行训练
+
+本教程简单演示单卡和多卡训练，详细的模型训练方法请参考[文档](../../docs/train/train_cn.md)。
+
+**单卡训练**
+
+在PaddleSeg目录下，执行如下命令，使用单卡GPU进行模型训练。
+
+```
+export CUDA_VISIBLE_DEVICES=0 # Linux下设置1张可用的卡
+# set CUDA_VISIBLE_DEVICES=0  # windows下设置1张可用的卡
+
+cd PaddleSeg
+
+python tools/train.py \
+       --config configs/smrt/pp_liteseg_stdc2.yml \
+       --do_eval \
+       --use_vdl \
+       --save_interval 1000 \
+       --save_dir output/pp_liteseg_stdc2
+```
+
+说明：
+* 上面脚本选择`pp_liteseg_stdc2`模型进行训练，所以加载`configs/smrt/pp_liteseg_stdc2.yml`配置文件，如果需要使用其他模型，可以修改`--config`输入参数。
+* `--do_eval`表示训练过程中会进行测试，`--save_interval`设置每训练多少轮会进行一次测试。
+* 训练结束后，精度最高的模型保存在`--save_dir`中，比如`output/pp_liteseg_stdc2/best_model/`。
+* 查看训练的log，可以知道模型训练的最高精度。
+* 为了实现最高的精度，大家可以适当调参，重点关注学习率lr、BatchSize、训练轮数iters、损失函数Loss等。
+
+
+**多卡训练**
+
+在PaddleSeg目录下，执行如下命令，使用多卡GPU进行模型训练。
+
+```
+export CUDA_VISIBLE_DEVICES=0,1,2,3 # 设置4张可用的卡
+
+python -m paddle.distributed.launch tools/train.py \
+       --config configs/smrt/pp_liteseg_stdc2.yml \
+       --do_eval \
+       --use_vdl \
+       --save_interval 1000 \
+       --save_dir output/pp_liteseg_stdc2
+```
+
+### 3.5 模型导出
+
+训练得到精度符合预期的模型后，可以导出预测模型，进行部署。详细的模型导出方法请参考[文档](../../docs/model_export_cn.md)。
+
+```
+python tools/export.py \
+       --config configs/smrt/pp_liteseg_stdc2.yml \
+       --model_path output/pp_liteseg_stdc2/best_model/model.pdparams \
+       --save_dir output/pp_liteseg_stdc2/infer_models
+```
+
+上面脚本加载`pp_liteseg_stdc2`模型精度最高的权重，导出预测模型保存在`output/pp_liteseg_stdc2/infer_models`目录。
+
+### 3.5 部署
+
+导出模型后，大家可以参考如下文档进行部署。
+
+| 端侧         | 库           | 教程   |
+| :----------- | :----------- | :----- |
+| 服务端端Python部署 | Paddle预测库 | [文档](../../docs/deployment/inference/python_inference_cn.md) |
+| 服务器端端C++部署 | PaddleInference预测库 | [文档](../../docs/deployment/inference/cpp_inference_cn.md) |
+| 移动端部署   | PaddleLite   | [文档](../../docs/deployment/lite/lite_cn.md) |
+| 前端部署     | PaddleJS     | [文档](../../docs/deployment/web/web_cn.md) |
+
+## 四、部署demo
+
+为了更方便大家部署，我们也提供了完备的可视化部署Demo，欢迎尝试使用
+
+* [Windows Demo下载地址](https://github.com/PaddlePaddle/PaddleX/tree/develop/deploy/cpp/docs/csharp_deploy)
+
+<div align="center">
+  <img src="https://user-images.githubusercontent.com/48433081/169064583-c931f4c0-dfd6-4bfa-85f1-be68eb351e4a.png"  width = "800" />  
+</div>
+
+* [Linux Demo下载地址](https://github.com/cjh3020889729/The-PaddleX-QT-Visualize-GUI)
+
+<div align="center">
+  <img src="https://user-images.githubusercontent.com/48433081/169065951-147f8d51-bf3e-4a28-9197-d717968de73f.png"  width = "800" />  
+</div>
+
+## 五、场景范例
+
+为了更方便大家更好的进行产业落地，PaddleSMRT也提供了详细详细的应用范例，欢迎大家使用。
+
+* 工业视觉
+  * [工业缺陷检测](https://aistudio.baidu.com/aistudio/projectdetail/2598319)
+  * [表计读数](https://aistudio.baidu.com/aistudio/projectdetail/2598327)
+  * [钢筋计数](https://aistudio.baidu.com/aistudio/projectdetail/2404188)
--- a/configs/smrt/base_cfg.yml
+++ b/configs/smrt/base_cfg.yml
+batch_size: 8
+iters: 10000
+
+train_dataset:
+  type: Dataset
+  dataset_root: data/defect_data
+  num_classes: 4
+  mode: train
+  train_path: data/defect_data/train.txt
+  transforms:
+    - type: ResizeStepScaling
+      min_scale_factor: 0.5
+      max_scale_factor: 2.0
+      scale_step_size: 0.25
+    - type: RandomPaddingCrop
+      crop_size: [512, 512]
+    - type: RandomHorizontalFlip
+    - type: RandomDistort
+      brightness_range: 0.4
+      contrast_range: 0.4
+      saturation_range: 0.4
+    - type: Normalize
+
+val_dataset:
+  type: Dataset
+  dataset_root: data/defect_data
+  num_classes: 4
+  mode: val
+  val_path: data/defect_data/val.txt
+  transforms:
+    - type: Normalize
+
+optimizer:
+  type: sgd
+  momentum: 0.9
+  weight_decay: 5.0e-4
+
+lr_scheduler:
+  type: PolynomialDecay
+  learning_rate: 0.01
+  end_lr: 0
+  power: 0.9
+  warmup_iters: 500
+  warmup_start_lr: 1.0e-5
\ No newline at end of file
--- a/configs/smrt/bisenetv2.yml
+++ b/configs/smrt/bisenetv2.yml
+_base_: './base_cfg.yml'
+
+model:
+  type: BiSeNetV2
+
+loss:
+  types:
+    - type: MixedLoss
+      losses:
+        - type: OhemCrossEntropyLoss
+          min_kept: 65000
+        - type: LovaszSoftmaxLoss
+      coef: [0.8, 0.2]
+  coef: [1, 1, 1, 1, 1]
\ No newline at end of file
--- a/configs/smrt/deeplabv3p_resnet50_os8.yml
+++ b/configs/smrt/deeplabv3p_resnet50_os8.yml
+_base_: './base_cfg.yml'
+
+model:
+  type: DeepLabV3P
+  backbone:
+    type: ResNet50_vd
+    output_stride: 8
+    multi_grid: [1, 2, 4]
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet50_vd_ssld_v2.tar.gz
+  backbone_indices: [0, 3]
+  aspp_ratios: [1, 12, 24, 36]
+  aspp_out_channels: 256
+  align_corners: False
+  pretrained: null
+
+
+loss:
+  types:
+    - type: MixedLoss
+      losses:
+        - type: OhemCrossEntropyLoss
+          min_kept: 65000
+        - type: LovaszSoftmaxLoss
+      coef: [0.8, 0.2]
+  coef: [1]
\ No newline at end of file