Skip to content

Crack your training task with Efficient Teacher

Jump to bottom
BowenXu edited this page Mar 15, 2023 · 1 revision

This article introduces how to quickly use Efficient Teacher to tackle your training tasks. The aproch is based on transfer learning and smi-supervised training of a pre-trained model on a large amount of richly annotated data. First let's review the normal transfer learning process: transfer

In contrast, here is the transfer learning process using Efficient Teacher for semi-supervised training:

transfer_ssod

We have provided the pre-trained YOLOv5l weights on Object365, which can be found at the following URL：https://github.com/AlibabaResearch/efficientteacher/releases/download/1.0/efficient-yolov5l-obj365.pt

The YAML configuration file for semi-supervised training can be found at the following URL：https://github.com/AlibabaResearch/efficientteacher/blob/main/configs/ssod/custom/yolov5l_transfer_ssod.yaml

Below is the hyperparameter definition for using pre-trained weights in a semi-supervised transfer learning training scheme, which is consistent with the standard YOLOv5 definition:

project: 'yolov5_ssod' #
adam: False
epochs: 30
weights: 'efficient-yolov5l-obj365.pt'
prune_finetune: False
linear_lr: True
# find_unused_parameters: True

hyp:
  lr0: 0.001
  hsv_h: 0.015
  hsv_s: 0.7
  hsv_v: 0.4
  lrf: 1.0
  scale: 0.9
  burn_epochs: 10
  no_aug_epochs: 0
  warmup_epochs: 3

The network structure definition and loss function definition are the same as in the standard YOLOv5:

Model:
  depth_multiple: 1.00  # model depth multiple
  width_multiple: 1.00  # 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
  anchor_t: 4.0

Dataset definition, where the more important parameter is sampler_type. Its function is to ensure that annotations of different classes appear evenly within a batch:

Dataset:
  data_name: 'coco'
  train: data/custom_train.txt  # 118287 images
  val: data/custom_val.txt  # 5000 images
  test: data/custom_val.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794^
  target: data/custom_unlabeled.txt
  nc: 2  # number of classes
  np: 0 #number of keypoints
  names: [ 'rick', 'morty']
  img_size: 640
  batch_size: 128
  sampler_type: 'class_balance'

Finally, the definition of the semi-supervised part. train_domain is the switch for the semi-supervised trainer. When set to False, the training script will automatically switch to supervised learning:

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
  ssod_hyp:
    with_gt: False
    mosaic: 1.0
    cutout: 0.5
    autoaugment: 0.5
    scale: 0.8
    degrees: 0.0
    shear: 0.0
Clone this wiki locally