PaddleSeg学习3——使用PP-LiteSeg模型对道路进行分割

1 准备环境

参考上一篇：Windows PaddleSeg c++部署

2 样本

车道线分割实战的智能车数据集

#标签
'背景'：0, 0.05
'单实线-黄': 1, 
'单实线-白': 2, 
'双实线-黄(包括一黄一白)': 3,
'双实线-白': 4, 
'单虚线-黄': 5, 
'单虚线-白': 6, 
'双虚线->    黄（包扩一黄一白）': 7, 
'双虚线-白': 8, 
'一实一虚-黄（包括一黄一白）': 9,
'一实一虚-白': 10, 
'多条车道线（线>2，不区分虚实与颜色）': 11, 
'实线减速车道线': 12, 
'虚线减速车道线': 13,
'锯齿线（可变车道线）': 14

3 准备config文件

参考下面两个文件
(1). 遥感道路分割的配置文件PaddleSeg-road/configs/road_seg/pp_liteseg_stdc1_deepglobe_1024x1024_80k.yml
(2). 车道线分割实战的配置文件
PaddleSeg/configs/ocrnet/ocrnet_hrnetw18_cityscapes_1024x512_160k_lovasz_softmax.yml

编写 pp_liteseg_stdc1_lane_1024x512.yml

batch_size: 8 
iters: 35000

optimizer:
  type: sgd
  momentum: 0.9
  weight_decay: 5.0e-4
  
lr_scheduler:
  type: PolynomialDecay
  learning_rate: 0.0025
  end_lr: 0
  power: 0.9
  warmup_iters: 1000
  warmup_start_lr: 1.0e-5
  
loss:
  types:
    - type: OhemCrossEntropyLoss
      min_kept: 260000
    - type: OhemCrossEntropyLoss
      min_kept: 260000
    - type: OhemCrossEntropyLoss
      min_kept: 260000
  coef: [1, 1, 1]
  
train_dataset:
  type: Dataset
  dataset_root: dataset/zhinengche
  num_classes: 15
  mode: train
  train_path: dataset/zhinengche/train_list.txt
  transforms:
    - type: ResizeStepScaling
      min_scale_factor: 0.5
      max_scale_factor: 2.0
      scale_step_size: 0.25
    - type: RandomRotation
      max_rotation: 90
    - type: RandomPaddingCrop
      crop_size: [768, 768]
    - type: RandomHorizontalFlip
    - type: RandomVerticalFlip
    - type: RandomDistort
      brightness_range: 0.3
      brightness_prob: 0.5
      contrast_range: 0.3
      contrast_prob: 0.5
      saturation_range: 0.3
      saturation_prob: 0.5
      hue_range: 15
      hue_prob: 0.5
    - type: Normalize

val_dataset:
  type: Dataset
  dataset_root: dataset/zhinengche/
  num_classes: 15
  mode: val
  val_path: dataset/zhinengche/val_list.txt
  transforms:
    - type: Normalize

model:
  type: PPLiteSeg
  backbone:
    type: STDC1
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/PP_STDCNet1.tar.gz
  arm_out_chs: [32, 64, 128]
  seg_head_inter_chs: [32, 64, 64]

4 模型训练

# train pp_liteseg
%cd ~/work/PaddleSeg-road/
!python train.py \
       --config configs/amy/pp_liteseg_stdc1_lane_1024x512.yml\
       #--do_eval \
       #--num_workers 3 \
       --save_interval 1000 \
       --save_dir output/pp_liteseg_stdc1_lane

5 模型预测

5.1 python预测

%cd ~/work/PaddleSeg-road
!python  predict.py \
    --config configs/amy/pp_liteseg_stdc1_lane_1024x512.yml \
    --model_path pp_liteseg_stdc1_lane/model.pdparams \
    --image_path Data/seg/test/5.jpg \
    --save_dir output/pp_liteseg_stdc1_lane

5.2 C++预测

（1）导出模型

%cd ~/work/PaddleSeg-road
!python  export.py \
    --config configs\amy\pp_liteseg_stdc1_lane_1024x512.yml \
    --model_path pp_liteseg_stdc1_lane/model.pdparams \
    --save_dir ./output/pp_liteseg_stdc1_lane

(2) c++预测

配置模型路径

修改输出的数据类型为 int64_t

5.3 预测结果展示