Enhanced version of original implementation from here. This is a detailed and simple enough implementation of YoloV3, and it is written in TensorFlow 2.0 ! We did those improvements compare to previous version:
- Add MobileNetV2 backbone, so it supports YoloV3 MobileNetV2;
- DCN network implemented;
- Add un-same-width image input (wider images);
- Add training on KITTI and nuScenes, also a TrafficLight detection model;
Welcome subscriber our ZhihuZhuanlan for more detailed Chinese tutorials! https://zhuanlan.zhihu.com/tensorflow2 .
there are several requirements for this repo:
absl-py
alfred-py
loguru
pytorch
If you got some package missing error, just install it.
After you prepared tfrecord file, you can kick off training easily:
INFO 05-09 16:28:23 train.py:51 - using YoloV3 model.
INFO 05-09 16:28:29 train.py:58 - loading dataset from: /media/jintain/wd/permenant/datasets/coco/*.tfrecord
INFO 05-09 16:28:33 train.py:113 - model resumed from: ./checkpoints/yolov3_coco-1.ckpt, start at epoch: 1
INFO 05-09 16:28:40 train.py:140 - Epoch: 1, iter: 0, total_loss: 1202.7585, pred_loss: [709.7193, 241.0539, 237.69351]
INFO 05-09 16:28:46 train.py:140 - Epoch: 1, iter: 10, total_loss: 886.7888, pred_loss: [368.6222, 253.66756, 249.7427]
INFO 05-09 16:28:52 train.py:140 - Epoch: 1, iter: 20, total_loss: 1315.7472, pred_loss: [639.7981, 431.42947, 229.50122]
INFO 05-09 16:28:58 train.py:140 - Epoch: 1, iter: 30, total_loss: 658.0850, pred_loss: [110.5262, 140.95578, 391.42163]
As you may notice, the training process if very shock in the start of training time. We will update some results after training.
We plan to do more things in the near future:
- Adding nms gpu in C++ and cuda;
- Accelerate the network using pruning and compressing model;
- Export to onnx and convert to ncnn.
this work original implemented by zzh8829, enhanced by jinfagang. authors got there right respectively.