EfficientDet_PyTorch

注意事项(NOTICE):

1. 训练请使用SGD优化器(with momentum 0.9). 不要使用Adam. 会造成不收敛.
   推荐lr: Pre_train(1e-4 -> 1e-3 -> 1e-2 -> 3e-2) -> Train(5e-2) -> Post_Train(3e-2 -> 1e-2 -> 1e-3)
   推荐weight_decay: 4e-5 或 1e-4 (区别不大)
   Use SGD optimizer for training(with momentum 0.9). Do not use Adam. It will cause a nonconvergence

2. 有两个分支，一个是按照论文书写(official)、一个是参考zylo117的代码(master)， 并使用了他的预训练模型书写(万分感谢)，请按实际情况选择
   There are two branches, one(official) was written according to the paper, the other(master) was written referring to the code of 'Zylo117' and use his pre-training model(thank you very much), please choose according to the actual situation

3. train_example.py 的意义是展示模型输入的格式
   The meaning of train_example.py is to show the format of the model input

4. 自己训练的时候，请使用EfficientNet预训练模型、并Freeze BackboneBN(推荐使用official)
   Use 'EfficientNet' pre-training model and 'Freeze BackboneBN' when you train yourself (Official is recommended)

Reference

1. 论文(paper):
   https://arxiv.org/pdf/1911.09070.pdf

2. 代码参考(reference code):
   https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch

3. EfficientNet 主干网代码来源(Backbone code source):
   https://github.com/Jintao-Huang/EfficientNet_PyTorch

4. 预训练模型来自(The pre-training model comes from):
   https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch
   因为修改了模型，所以我把预训练模型的state_dict进行了重组，并进行发布
   (Because I changed the model, I reorganized the state_dict for the pretraining model and release it)

权重见 release. 或在百度云中下载:
链接：https://pan.baidu.com/s/1VrO0eBmSHlB8_haEJ7WbuA
提取码：2kq9

使用方式(How to use)

1. 预测图片(Predict images)

python3 pred_image.py

2. 预测视频(Predict video)

python3 pred_video.py

3. 简单的训练案例(Simple training cases)

python3 train_example.py

4. 训练

python3 make_dataset.py
python3 train.py

性能

如果打不开可在images/与docs/文件夹中查看

d0效果

运行环境(environment)

torch 1.2.0
torchvision 0.4.0