如何训练LightWeight OpenPose

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

解压至文件夹

２　安装要求的环境　　pip install -r requirements.txt

## Training

训练

Training consists of 3 steps (given AP values for full validation dataset):

训练由三个步骤组成(给定全验证集的AP值)

* Training from MobileNet weights. Expected AP after this step is ~38%.

从MobileNet 权重开始训练,经过此步骤期待AP约为38%

* Training from weights, obtained from previous step. Expected AP after this step is ~39%.

从上一步骤得到的权重训练,经过此步骤期待AP约为39%

* Training from weights, obtained from previous step and increased number of refinement stages to 3 in network. Expected AP after this step is ~40% (for the network with 1 refinement stage, two next are discarded).

训练上一步骤得到的权重,增加改善层到3.经过此步骤,期待AP约为40%

1. Download pre-trained MobileNet v1 weights `mobilenet_sgd_68.848.pth.tar` from: [https://github.com/marvis/pytorch-mobilenet](https://github.com/marvis/pytorch-mobilenet) (sgd option). If this doesn't work, download from [GoogleDrive](https://drive.google.com/file/d/18Ya27IAhILvBHqV_tDp0QjDFvsNNy-hv/view?usp=sharing).

下载预训练的MobileNet V1 权重mobilenet_sgd_68.848.pth.tar,地址为[https://github.com/marvis/pytorch-mobilenet](https://github.com/marvis/pytorch-mobilenet) (sgd option). If this doesn't work, download from [GoogleDrive](https://drive.google.com/file/d/18Ya27IAhILvBHqV_tDp0QjDFvsNNy-hv/view?usp=sharing).

2. Convert train annotations in internal format. Run `python scripts/prepare_train_labels.py --labels /annotations/person_keypoints_train2017.json`. It will produce `prepared_train_annotation.pkl` with converted in internal format annotations.

转换训练的注释框到内部格式,运行python scripts/prepare_train_labels.py --labels /annotations/person_keypoints_train2017.json 会产生prepared_train_annotation.pkl文件.

[OPTIONAL] For fast validation it is recommended to make *subset* of validation dataset. Run `python scripts/make_val_subset.py --labels /annotations/person_keypoints_val2017.json`. It will produce `val_subset.json` with annotations just for 250 random images (out of 5000).

3. To train from MobileNet weights, run `python train.py --train-images-folder /train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder /val2017/ --checkpoint-path /mobilenet_sgd_68.848.pth.tar --from-mobilenet`

从MObileNetV1 权重训练,运行python train.py --train-images-folder /train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder /val2017/ --checkpoint-path /mobilenet_sgd_68.848.pth.tar --from-mobilenet

4. Next, to train from checkpoint from previous step, run `python train.py --train-images-folder /train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder /val2017/ --checkpoint-path /checkpoint_iter_420000.pth --weights-only`

5. Finally, to train from checkpoint from previous step and 3 refinement stages in network, run `python train.py --train-images-folder /train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder /val2017/ --checkpoint-path /checkpoint_iter_280000.pth --weights-only --num-refinement-stages 3`. We took checkpoint after 370000 iterations as the final one.

最终,为了从上一步骤和3 网络调优阶段的断点进行训练,运行run `python train.py --train-images-folder /train2017/ --prepared-train-labels prepared_train_annotation.pkl --val-labels val_subset.json --val-images-folder

We did not perform the best checkpoint selection at any step, so similar result may be achieved after less number of iterations.
我们不在任何步骤选择最好的断点,迭代次数少了会导致相似的结果.