becauseofai / efficientteacher Goto Github PK

Before we proceed with semi-supervised training, we need you to convert your own model trained with the YOLOv5 open source framework into a format that we can recognize. If you are using version 6.0 or later, the process is very fast and can be completed in five minutes:

1. Convert Model

• First, you need to write a yaml file. You can directly modify the file configs/custom/yolov5l_custom.yaml. If your model is yolov5l, then you only need to modify the nc in Dataset to the number of your detection classes and then modify the names. If your model is other depth and width configurations, then modify the depth_multiple and width_multiple to the corresponding configurations.
• After having this yaml file, go to the scripts folder to modify the convert_pt_to_efficient.py, fill in your business model pt, yaml, and the exported pt file.
• OK, you have already converted your pt file into a version that our algorithm library can recognize. Don't worry, if you need to export the model to onnx or export it back to your own YOLOv5 algorithm library, we also provide corresponding scripts.

2. Validation

• This step is to verify that the converted model still maintains the corresponding accuracy and recall on your validation set, so we hope you use the following script to verify the current model again, so that you also have a semi-supervised object detection baseline, in order to confirm that Efficient Teacher really works on your dataset.

• modify the val: data/custom_val.txt in yolov5l_custom.yaml into your own validation set path, then run the following code:

  python val.py --cfg configs/sup/custom/yolov5l_custom.yaml --weights efficient-yolov5l.pt 
  

• if you modify the file read path, you will need to add the corresponding modifications in the utils/dataloader.py file, just like you did when you modified the code for YOLOv5.

3. Supervised Training(Optional)

• modify the train: data/custom_train.txt in yolov5l_custom.yaml, and then enter the following script.

  export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
  python -m torch.distributed.launch --nproc_per_node 8 --master_addr 127.0.0.2 --master_port 29502 train.py --cfg configs/sup/custom/yolov5l_custom.yaml 
  

4. SSOD Training

• modify the train: data/custom_train.txt in yolov5l_custom.yaml, creat an unlabeled data image list using following command:

  find <unlabel_path> -name "*.jpg" >> unlabel.txt
  

• change the target: data_custom_target.txt to target:unlabel.txt in yolov5l_custom.yaml, and psate the following config into yolov5l_custom.yaml:

  SSOD:
    train_domain: True
    nms_conf_thres: 0.1
    nms_iou_thres: 0.65
    teacher_loss_weight: 1.0
    cls_loss_weight: 0.3
    box_loss_weight: 0.05
    obj_loss_weight: 0.7
    loss_type: 'ComputeStudentMatchLoss'
    ignore_thres_low: 0.1
    ignore_thres_high: 0.6
    uncertain_aug: True
    use_ota: False
    multi_label: False
    ignore_obj: False
    pseudo_label_with_obj: True
    pseudo_label_with_bbox: True
    pseudo_label_with_cls: False
    with_da_loss: False
    da_loss_weights: 0.01
    epoch_adaptor: True
    resample_high_percent: 0.25
    resample_low_percent: 0.99
    ema_rate: 0.999
    cosine_ema: True
    imitate_teacher: False
    ssod_hyp:
      with_gt: False
      mosaic: 1.0
      cutout: 0.5
      autoaugment: 0.5
      scale: 0.8
      degrees: 0.0
      shear: 0.0
  

• you have now completed a re-write yaml file for a SSOD training, enter the following script:

  export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
  python -m torch.distributed.launch --nproc_per_node 8 --master_addr 127.0.0.2 --master_port 29502 train.py --cfg configs/sup/custom/yolov5l_custom.yaml 
  

1. Edit your semi-supervised training task according to configs/ssod/custom/yolov5l_custom_ssod.yaml, where the train/val/test in the Dataset should be filled in according to your project's original txt, and the target needs you to index an unlabeled dataset that you expect. We suggest using the images of the COCO train dataset, and a unlabeled dataset can be generated by entering "find img_dir -name "*.jpg" >> target_img.txt".

2. rewrite nc and names in yaml according to your individual tasks.

3. train SSOD model from scratch, the default setting is to first conduct a 220-epoch burn-in training, and then enter the SSOD training, which has been introduced in our paper.

   export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
   python -m torch.distributed.launch --nproc_per_node 8 --master_addr 127.0.0.2 --master_port 29502 train.py --cfg configs/ssod/custom/yolov5l_custom_ssod.yaml 
   

4. validation SSOD model

   python val.py --cfg configs/ssod/custom/yolov5l_custom_ssod.yaml --weights ssod-yolov5l.pt  --val-ssod