前期准备：
1 下载ＣＯＣＯ２０１７数据集，数据集地址为：http://cocodataset.org/#download](http://cocodataset.org/#download
解压至<COCO_HOME>文件夹
２ 安装要求的环境 pip install -r requirements.txt
训练
训练包含三个步骤
１．训练MobileNet权重，经过此轮训练期望ＡＰ值达到３８％
２．训练上一阶段的权重，期待达到ＡＰ３９％
３训练上一阶段的权重，期待达到ＡＰ４０％
具体步骤：
１．下载MobileNet v1权重 mobilenet_sgd_68.848.pth.tar
２． 转换标签格式
python scripts/prepare_train_labels.py --labels <COCO_HOME>/annotations/person_keypoints_train2017.json
产生prepared_train_annotation.pkl
３．python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/mobilenet_sgd_68.848.pth.tar --from-mobilenet
4. python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/checkpoint_iter_420000.pth --weights-only
5. 继续训练，三步调优, python train.py --train-images-folder <COCO_HOME>/train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder <COCO_HOME>/val2017/ --checkpoint-path <path_to>/checkpoint_iter_280000.pth --weights-only --num-refinement-stages 3 采用 checkpoint 370000 迭代后作为最终的结果．
We did not perform the best checkpoint selection at any step, so similar result may be achieved after less number of iterations.
