/ClassificationWork

Training common classification task

Primary LanguagePython

Training code for Common Classification task

​ This repo is used to train the common classification model with many backbones and tricks (algorithm) for improve the model's performance. I implement it with pytorch.

  • Data Augmentation: mixup, cut-mix, random crop, flip, random hsv, random brightness and so on.
  • Some Algorithm: mean-teacher, knowledge-distilling, warm-up training
  • backbone: resnet, mobilenetv2, mobilenetv3, resnet10, chostnet, efficientnet,efficientv2,RepVgg, shufflenetv1/v2. Some models can be create by the pretrained models in torchbision

Requirments

  • torch >=1.7.1+cu101
  • python3.7
  • numpy>=1.12.1
  • scipy>=0.19.0
  • Cython
  • pycocotools
  • lap
  • motmetrics
  • opencv-python
  • funcy

Usage

  1. Prepare Data
  • Firstly,put your dataset to the DataSet dir. Different class with different dir . Here is a demo:
DataSet/maskdata, you refer this dir structure

pos(face with mask)

neg(face without mask)

  • Secondly, use the train_val_caffe.py to generate the label list file. Modify the line50-54 according your dataset

    # classes name list
classNames=["neg","pos"]
cat_ids = {v: i for i, v in enumerate(classNames)}
#class data root in Dataset
train_root="maskdata/train"
test_root="maskdata/val"

  1. Training

    ​ You can choose different .py to train your model. For common training method, you can choose the common_train.py , Just modify the params in the argparse for your task and data. And there are other methods such as:

    • knowledge_distiliing_training.py,

    • mean_teacher_training.py

    • mean_teacher_training_resnet10.py

  2. export to onnx

    please refer the export_onnx.py, and modify some params(argparse) according to your model