tidalove/yolox

1 2# Updates notes3 4## 【2021/08/19】5 6* Support image caching for faster training, which requires large system RAM. 7* Remove the dependence of apex and support torch amp training. 8* Optimize the preprocessing for faster training 9* Replace the older distort augmentation with new HSV aug for faster training and better performance. 10 11### 2X Faster training12 13We optimize the data preprocess and support image caching with `--cache` flag:14 15```shell16python tools/train.py -n yolox-s -d 8 -b 64 --fp16 -o [--cache]17                         yolox-m18                         yolox-l19                         yolox-x20```21* -d: number of gpu devices22* -b: total batch size, the recommended number for -b is num-gpu * 823* --fp16: mixed precision training24* --cache: caching imgs into RAM to accelarate training, which need large system RAM.25 26### Higher performance27 28New models achieve **~1%** higher performance! See [Model_Zoo](model_zoo.md) for more details.29 30### Support torch amp31 32We now support torch.cuda.amp training and Apex is not used anymore.33 34### Breaking changes35 36We remove the normalization operation like -mean/std. This will make the old weights **incompatible**.37 38If you still want to use old weights, you can add `--legacy' in demo and eval:39 40```shell41python tools/demo.py image -n yolox-s -c /path/to/your/yolox_s.pth --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 640 --save_result --device [cpu/gpu] [--legacy]42```43 44and 45 46```shell47python tools/eval.py -n  yolox-s -c yolox_s.pth -b 64 -d 8 --conf 0.001 [--fp16] [--fuse] [--legacy]48                         yolox-m49                         yolox-l50                         yolox-x51```52 53But for deployment demo, we don't support the old weights anymore. Users could checkout to YOLOX version 0.1.0 to use legacy weights for deployment54 55 56