This repository is forked from https://github.com/eriklindernoren/PyTorch-YOLOv3, I modify it to support multi GPU train.
If you want to use multi GPU train, you should modify these below:
- pad targets in utils/datasets.py -> collate_fn() function line149
- remove pad targets in utils/utils.py -> build_targets() function line 301
- re-calculate batch-index(targets) in utils/utils.py -> build_targets() function line 325
- move model outputs/loss from cpu to gpu in models.py -> Darknet.forward() function line 330
- calculate loss sum from all GPUs in train.py -> line 121
- modify log info in train.py -> line 130
关于为何做此修改？可以参考https://github.com/ujsyehao/yolov3-multigpu/issues/8 大家跑之前一定要修改imgs_per_gpu在build_targets()函数中
- mAP计算 论文中是55.3, 原来仓库使用conf_thresh=0.001@IOU=0.5 mAP为53.6
- 修改datasets.py中collate_fn(), 具体原因可查看eriklindernoren/PyTorch-YOLOv3#243 -> mAP为54.9
- 修改utils.py中get_batch_statistics(), 具体原因可查看eriklindernoren/PyTorch-YOLOv3#233 -> mAP为52.6
- 原来仓库作者提供的coco2014 val 5k dataset和coco官网提供的5k val dataset不一致
-
gt box越界错误 fix gt box may cross boundary in utils/utils.py -> build_targets() function line 332
-
多卡读取weights
４．多卡log信息
- data augmentation
- multi scale train √
- multi GPU train √
- label smooth
- mix up
- cosine lr √
- warmup √
- group normalization (deprecated)
- focal loss
- soft nms (not supported)
- GIOU