Writing in the front
This is my first technical blog post and the beginning of what I hope will be a good habit of documenting and summarizing projects before the start of my graduate career. To be clear, the goal of this blog is to make deep learning easier for more people, so it’s also a summary of the pitfalls.
Know your hardware
At the beginning of model training, you must first understand your hardware device. The purpose of this project is to deploy the trained model on the eageboardFz3B board, so before this, you must first understand the version and dependency of the model that your hardware device can bear. The line is summarized as follows:
| Eageboard dependent version number | |
|---|---|
| paddlepaddle | < = 1.8.5 |
| PaddleDetection | Realse 0.5 |
| PaddleSlim | = 2.9.0 |
The software platform
Baidu Flying slurry training platform AIstudio. The project address
Model training
Because of the requirements of the project, we have to train our own data set, not the open data set. Of course, if your project can find an alternative data set on the Internet, you can also choose the open data set, so you don’t need to annotate the data.
1. Prepare data sets
The annotation of data set is mainly about the annotation of VOC data set. Since I finally choose to use the data set in VOC format, the content of converting Labelme into COCO format is omitted. If necessary, please refer to [help document](How to train custom data set — PaddleDetection 0.1 document).
Tool use
Use Sprite annotation Assistant to complete annotation tasks,
2. Use aiStudio platform for training
2.1 Creating a Project
AI Studio provides scripts and Jupyter Notebooks; For example, ssD_mobilenet model needs to be deployed to the terminal, so it is necessary to know the computing power of your terminal in advance. This is particularly important in practice, because the previous work did not do a good job in the investigation, leading to the training of yolov3 model because the model was too large and the board could not provide the required computing power, so the expected effect could not be achieved. Finally, only the lightweight model can be used to complete the detection task. The performance of common models on edgeBoard is shown in the following table.
! Unzip PaddleDetection - release - 0.5. Zip
Extract data set
! unzip data/data130283/trytwo3.zip -d work/
Partition data set
import random
importTXT and val.txt random. Seed (2020)
xml_dir = '/home/aistudio/work/trytwo/ANNOTATIONS'# tag file address img_dir ='/home/aistudio/work/trytwo/IMAGES'Path_list = list()for img in os.listdir(img_dir):
img_path = os.path.join(img_dir,img)
xml_path = os.path.join(xml_dir,img.replace('png'.'xml'))
path_list.append((img_path, xml_path))
random.shuffle(path_list)
ratio = 0.9
train_f = open('/home/aistudio/work/train.txt'.'w'Val_f = open()'/home/aistudio/work/val.txt' ,'w')# generate validation filefor i ,content in enumerate(path_list):
img, xml = content
text = img + ' ' + xml + '\n'
if i < len(path_list) * ratio:
train_f.write(text)
else: val_F.write (text) train_F.close () val_F.close ()'Slope'.'Numone'.'Freeezone'.'NumTwo'.'STOP'.'CASD'.'GasStation'.'GrossWalk'.'Cone']# set the category you want to detectwith open('/home/aistudio/work/label_list.txt'.'w') as f:
for text in label:
f.write(text+'\n')
Copy the codeParameter configuration
Set the path configuration parameters for your data set under the path paddledisage-release -0.5/configs/ SSD/ssd_mobilenet_v1_VOC. yml. Here are the parameters you may need to change. num_classes
TrainReader:
inputs_def:
image_shape: [3.300.300]
fields: ['image'.'gt_bbox'.'gt_class']
dataset:
!VOCDataSet
anno_path: train.txt #dataset_dir: /home/Aistudio /work # needs to be modifieduse_default_label: false# Default to True, change tofalse, the same belowsample_transforms:
EvalReader:
inputs_def:
image_shape: [3.300.300]
fields: ['image'.'gt_bbox'.'gt_class'.'im_shape'.'im_id'.'is_difficult']
dataset:
!VOCDataSet
anno_path: val.txt #dataset_dir: /home/Aistudio/work # modifieduse_default_label: false
sample_transforms:
TestReader:
inputs_def:
image_shape: [3.300.300]
fields: ['image'.'im_id'.'im_shape']
dataset:
!ImageFolder
anno_path: /home/Aistudio/work/label_list. TXT # changeuse_default_label: false
sample_transforms:
Copy the codeStart training
! Python PaddleDetection – release – 0.5 / tools/train. Py – cPaddleDetection – release – 0.5 / configs/SSD/ssd_mobilenet_v1_voc yml Seeing such information is considered a successful beginning of training
At the end of the training, it will automatically generate an Output folder, and then we can predict an image and see what it looks like and then you can modify it for your own needs
Model to predict
python PaddleDetection-release-0.5/tools/infer.py \
-c PaddleDetection-release-0.5/configs/ssd/ssd_mobilenet_v1_voc.yml \
-o weights=output/ssd_mobilenet_v1_voc/best_model.pdparams \
--infer_img=work/trytwo/IMAGES/1010.png
Copy the codeThe export model
python PaddleDetection-release-0.5/tools/export_model.py -c PaddleDetection-release-0.5/configs/ssd/ssd_mobilenet_v1_voc.yml --output_dir=./inference_model -o weights=output/ssd_mobilenet_v1_voc/best_model.pdparams TestReader.inputs_def.image_shape=[3.300.300]
Copy the codePaddleDemo is required to deploy the model, just follow the JSON file sample configuration, and you have completed all the steps that can be done on the AIStudio platform.
Video to predict
python PaddleDetection-release-0.5/deploy/python/infer.py --model_dir=inference_model/ssd_mobilenet_v1_voc --video_file=testes.mp4 --thresh=0.8
Copy the codeThe results of
Model deployment and tuning
You can see that the model does not always behave as you would expect when deployed to an end device, so this requires further tuning work, some parameter tuning reference
