使用多 GPU 训练 pytorch-yolov3.zip

# yolov3-multigpu ## train pytorch-yolov3 with multi GPU 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 ## 原来仓库问题修复: 1. 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https://github.com/eriklindernoren/PyTorch-YOLOv3/issues/243 -> mAP为54.9 * 修改utils.py中get_batch_statistics(), 具体原因可查看https://github.com/eriklindernoren/PyTorch-YOLOv3/issues/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 2. gt box越界错误 fix gt box may cross boundary in utils/utils.py -> build_targets() function line 332 3. 多卡读取weights ４．多卡log信息 ## TODO tricks: ### training tricks * data augmentation * multi scale train √ * multi GPU train √ * label smooth * mix up * cosine lr √ * warmup √ * group normalization (deprecated) ### detection tricks * focal loss * soft nms (not supported) * GIOU