paddlers/tutorials/train/README_en.md

Tutorial - Training Model

Sample code using the PaddleRS training model is curated in this directory. The code provides automatic downloading of sample data, and uses GPU to train the model.

Sample code path	Task	Model
change_detection/bit.py	Change Detection	BIT
change_detection/cdnet.py	Change Detection	CDNet
change_detection/changeformer.py	Change Detection	ChangeFormer
change_detection/dsamnet.py	Change Detection	DSAMNet
change_detection/dsifn.py	Change Detection	DSIFN
change_detection/fc_ef.py	Change Detection	FC-EF
change_detection/fc_siam_conc.py	Change Detection	FC-Siam-conc
change_detection/fc_siam_diff.py	Change Detection	FC-Siam-diff
change_detection/fccdn.py	Change Detection	FCCDN
change_detection/p2v.py	Change Detection	P2V-CD
change_detection/snunet.py	Change Detection	SNUNet
change_detection/stanet.py	Change Detection	STANet
classification/condensenetv2.py	Scene Classification	CondenseNet V2
classification/hrnet.py	Scene Classification	HRNet
classification/mobilenetv3.py	Scene Classification	MobileNetV3
classification/resnet50_vd.py	Scene Classification	ResNet50-vd
image_restoration/drn.py	Image Restoration	DRN
image_restoration/esrgan.py	Image Restoration	ESRGAN
image_restoration/lesrcnn.py	Image Restoration	LESRCNN
object_detection/faster_rcnn.py	Object Detection	Faster R-CNN
object_detection/ppyolo.py	Object Detection	PP-YOLO
object_detection/ppyolo_tiny.py	Object Detection	PP-YOLO Tiny
object_detection/ppyolov2.py	Object Detection	PP-YOLOv2
object_detection/yolov3.py	Object Detection	YOLOv3
semantic_segmentation/bisenetv2.py	Image Segmentation	BiSeNet V2
semantic_segmentation/deeplabv3p.py	Image Segmentation	DeepLab V3+
semantic_segmentation/factseg.py	Image Segmentation	FactSeg
semantic_segmentation/farseg.py	Image Segmentation	FarSeg
semantic_segmentation/fast_scnn.py	Image Segmentation	Fast-SCNN
semantic_segmentation/hrnet.py	Image Segmentation	HRNet
semantic_segmentation/unet.py	Image Segmentation	UNet

Environmental Preparation

• PaddlePaddle installation

  • Version requirements: PaddlePaddle>=2.2.0

• PaddleRS installation

The PaddleRS code will be updated as the development progresses. You can install the develop branch to use the latest features as follows:

git clone https://github.com/PaddlePaddle/PaddleRS
cd PaddleRS
git checkout develop
pip install -r requirements.txt
python setup.py install

If the downloading of dependencies is slow or times out when using python setup.py install, you can create setup.cfg in the same directory as setup.py and with the following content, then the download can be accelerated through Tsinghua source:

[easy_install]
index-url=https://pypi.tuna.tsinghua.edu.cn/simple

• (Optional) GDAL installation

PaddleRS supports reading of various types of satellite data. To use the full data reading functionality of PaddleRS, you need to install GDAL as follows:

• Linux / MacOS

conda is recommended for installation:

conda install gdal

• Windows

Windows users can download GDAL wheels from this site. Please choose the wheel according to the Python version and the platform. Take GDAL‑3.3.3‑cp39‑cp39‑win_amd64.whl as an example, run the following command to install:

pip install GDAL‑3.3.3‑cp39‑cp39‑win_amd64.whl

*Docker Installation

1. Pull from dockerhub:

   docker pull paddlepaddle/paddlers:1.0.0
   

• (Optional) Build from scratch. Select the base image for PaddlePaddle by setting PPTAG. You can build the image in a CPU-only environment or in GPU environments.

  git clone https://github.com/PaddlePaddle/PaddleRS
  cd PaddleRS
  docker build -t <imageName> .  # default is 2.4.1-cpu version
  # docker build -t <imageName> . --build-arg PPTAG=2.4.1-gpu-cuda10.2-cudnn7.6-trt7.0  # One of the gpu versions of 2.4.1
  # Other Tag refer to: https://hub.docker.com/r/paddlepaddle/paddle/tags
  

1. Start a container

   docker images  # View the ID of an image
   docker run -it <imageID>
   

Start Training

• After PaddleRS is installed, run the following commands to launch training with a single GPU. The script will automatically download the training data. Take DeepLab V3+ image segmentation model as an example:

  # Specifies the GPU device number to be used
  export CUDA_VISIBLE_DEVICES=0
  python tutorials/train/semantic_segmentation/deeplabv3p.py
  

• If multiple GPUs are required for training, for example, two graphics cards, run the following command:

  python -m paddle.distributed.launch --gpus 0,1 tutorials/train/semantic_segmentation/deeplabv3p.py
  

VisualDL Visual Training Metrics

Set the use_vdl argument passed to the train() method to True, and then the training log will be automatically saved in VisualDL format in a subdirectory named vdl_log under the directory specified by save_dir(a user-specified path) during the model training process. You can run the following command to start the VisualDL service and view the indicators and metrics. We also take DeepLab V3+ as an example:

# The specified port number is 8001
visualdl --logdir output/deeplabv3p/vdl_log --port 8001

Once the service is started, open https://0.0.0.0:8001 or https://localhost:8001 in your browser to access the VisualDL page.