阿里最新半监督学习框架——EfficientTeacher使用指南

源码地址

https://github.com/AlibabaResearch/efficientteacher

创建虚拟环境

conda create -n SSOD python=3.11 -y
conda activate SSOD

安装依赖

pip install -r requirements.txt

数据集划分

这里提供一种思路：具体的数据集划分按照自己的需求进行，假设现在有1000张图片数据用于训练。
1、首先划分出全监督的数据集：取出10%作为全监督训练的数据，即100张。在100张的基础上划分出train、val、test，按照7:2:1的比例划分。在YOLOv5模型基础上训练出一个pt文件，命名为：pre_v5.pt

Dataset:
  data_name: 'crowdhuman'
  train: dataset_crowdhuman/train.txt  # 70 images
  val: dataset_crowdhuman/val.txt  # 20 images
  test: dataset_crowdhuman/test.txt # 10 images
  nc: 2  # 类别数量
  np: 0 # number of keypoints
  names: ['head', 'person']
  img_size: 640
  batch_size: 64
  workers: 8

2、进行半监督学习的训练，拿出剩余的90%数据参与到新的训练中，该数据没有标签，只有图片。

Dataset:
  data_name: 'crowdhuman'
  train: dataset_crowdhuman/train.txt  # 70 images
  val: dataset_crowdhuman/val.txt  # 20 images
  test: dataset_crowdhuman/test.txt # 10 images
  target: dataset_crowdhuman/unlabeled_train.txt # 900 images
  nc: 2  # number of classes
  np: 0 # number of keypoints
  names: ['head', 'person']
  img_size: 640
  batch_size: 32
  workers: 8

在这里给出两个yaml文件，分别用于全监督和半监督学习。

# 全监督学习的YAML文件
# Parameters
project: 'sup_runs'
name: 'yolov5n_crowdhuman_1_percent'
adam: False
epochs: 500
weights: 'efficient-yolov5n.pt'	# 初始化权重用YOLOv5官方的
prune_finetune: False
linear_lr: True
cache: True # Requires large memory
hyp:
  lr0: 0.01
  lrf: 0.01
  hsv_h: 0.015
  hsv_s: 0.7
  hsv_v: 0.4
  scale: 0.9
  translate: 0.1
  fliplr: 0.5
  mosaic: 1.0
  mixup: 0.1
  warmup_epochs: 3
  no_aug_epochs: 30

Model:
  depth_multiple: 0.33  # model depth multiple
  width_multiple: 0.25  # layer channel multiple
  Backbone: 
    name: 'YoloV5'
    activation: 'SiLU'
  Neck: 
    name: 'YoloV5' 
    in_channels: [256, 512, 1024]
    out_channels: [256, 512, 1024]
    activation: 'SiLU'
  Head: 
    name: 'YoloV5'
    activation: 'SiLU'
  anchors: [[10,13, 16,30, 33,23],[30,61, 62,45, 59,119],[116,90, 156,198, 373,326]]  # P5/32]
Loss:
  type: 'ComputeLoss'
  cls: 0.3
  obj: 0.7
  box: 0.05
  anchor_t: 4.0

Dataset:
  data_name: 'crowdhuman'
  train: dataset_crowdhuman/train.txt  # 70 images
  val: dataset_crowdhuman/val.txt  # 20 images
  test: dataset_crowdhuman/test.txt # 10 images
  nc: 2  # number of classes
  np: 0 # number of keypoints
  names: ['head', 'person']
  img_size: 640
  batch_size: 64
  workers: 8

# 半监督学习的YAML文件
# Parameters
project: 'ssod_runs'
name: 'yolov5n_crowdhuamn_1per_exp1_ssod'
adam: False
epochs: 300
weights: 'sup_runs/yolov5n_crowdhuman_1_percent/weights/best.pt'
prune_finetune: False
linear_lr: True
cache: True

hyp:
  lr0: 0.01
  lrf: 1.0
  hsv_h: 0.015
  hsv_s: 0.7
  hsv_v: 0.4
  scale: 0.9
  translate: 0.1
  fliplr: 0.5
  mosaic: 1.0
  mixup: 0.1
  warmup_epochs: 0
  no_aug_epochs: 0
  burn_epochs: 0

Model:
  depth_multiple: 0.33  # model depth multiple
  width_multiple: 0.25  # layer channel multiple
  Backbone: 
    name: 'YoloV5'
    activation: 'SiLU'
  Neck: 
    name: 'YoloV5' 
    in_channels: [256, 512, 1024]
    out_channels: [256, 512, 1024]
    activation: 'SiLU'
  Head: 
    name: 'YoloV5'
    activation: 'SiLU'
  anchors: [[10,13, 16,30, 33,23],[30,61, 62,45, 59,119],[116,90, 156,198, 373,326]]  # P5/32]
Loss:
  type: 'ComputeLoss'
  cls: 0.3
  obj: 0.7
  box: 0.05
  anchor_t: 4.0

Dataset:
  data_name: 'crowdhuman'
  train: dataset_crowdhuman/train.txt  # 70 images
  val: dataset_crowdhuman/val.txt  # 20 images
  test: dataset_crowdhuman/test.txt # 10 images
  target: dataset_crowdhuman/unlabeled_1_percent.txt # 900 images
  nc: 2  # number of classes
  np: 0 # number of keypoints
  names: ['head', 'person']
  img_size: 640
  batch_size: 32
  workers: 8

SSOD:
  train_domain: True
  nms_conf_thres: 0.1
  nms_iou_thres: 0.7
  teacher_loss_weight: 0.5
  cls_loss_weight: 0.3
  box_loss_weight: 0.05
  obj_loss_weight: 0.7
  loss_type: 'ComputeStudentMatchLoss'
  ignore_thres_low: 0.4
  ignore_thres_high: 0.45
  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
  # dynamic_thres: True
  ssod_hyp:
    with_gt: False
    mosaic: 1.0
    cutout: 0.5
    autoaugment: 0.5
    scale: 0.8
    degrees: 0.0
    shear: 0.0

训练自己的数据集

#### 1. 转换权重. (yolo => efficient)
python scripts/mula_convertor/convert_pt_to_efficient.py

#### 2. 有监督训练
python train.py --cfg configs/sup/custom/yolov5n_1per.yaml

#### 3. 半监督训练
python train.py --cfg configs/ssod/custom/yolov5n_1per_ssod_exp1.yaml

#### 4. 测试全监督模型.
python val.py --data dataset_crowdhuman/data.yaml --weights sup_runs/yolov5n_crowdhuman_1_percent/weights/best.pt --task test --name yolov5n_crowdhuman_1_percent

#### 5. 测试半监督模型.
python val.py --cfg configs/ssod/custom/yolov5n_1per_ssod_exp1.yaml --weights ssod_runs/yolov5n_crowdhuamn_1per_exp1_ssod/weights/best.pt --task test --name yolov5n_crowdhuamn_1per_exp1_ssod --val-ssod

#### 6. 推理.
python detect.py --weights sup_runs/yolov5n_crowdhuman_1_percent/weights/best.pt --source dataset_crowdhuman/images/test/284193,bf20000e03d9239.jpg --name yolov5n_crowdhuman_1_percent  

#### 7. 转换权重. (efficient => yolo)
python scripts/mula_convertor/convert_pt_to_efficient.py