基于飞桨实现乒乓球时序动作定位大赛-baseline

| - data
	| - data123004
		| - Features_competition_test_A.tar.gz
	| - data122998
		| - Features_competition_train.tar.gz
		| - label_cls14_train.json

{'image_feature': array([[-0.00178786, -0.00247065,  0.00754537, ..., -0.00248864,
        -0.00233971,  0.00536158],
       [-0.00212389, -0.00323782,  0.0198264 , ...,  0.00029546,
        -0.00265382,  0.01696528],
       [-0.00230571, -0.00363361,  0.01017699, ...,  0.00989012,
        -0.00283369,  0.01878656],
       ...,
       [-0.00126995,  0.01113492, -0.00036558, ...,  0.00343453,
        -0.00191288, -0.00117079],
       [-0.00129959,  0.01329842,  0.00051888, ...,  0.01843636,
        -0.00191984, -0.00067066],
       [-0.00134973,  0.02784026, -0.00212213, ...,  0.05027904,
        -0.00198008, -0.00054018]], dtype=float32)}
<class 'numpy.ndarray'>

# label_cls14_train.json
{
    'fps': 25,    #视频帧率
    'gts': [
        {
            'url': 'name_of_clip.mp4',      #名称
            'total_frames': 6341,    #总帧数（这里总帧数不代表实际视频帧数，以特征数据集维度信息为准）
            'actions': [
                {
                    "label_ids": [7],   #动作类型编号
                    "label_names": ["name_of_action"],     #动作类型
                    "start_id": 201,  #动作起始时间,单位为秒
                    "end_id": 111    #动作结束时间,单位为秒
                },
                ...
            ]
        },
        ...
    ]
}

# 检查数据集所在路径!tree -L 3 /home/aistudio/data

%cd /home/aistudio/data# 创建特征数据集存放目录%mkdir Features_train Features_test# 分别解压缩训练和测试特征数据到各自目录下，解压完成后删除原压缩文件，节省空间!tar -xf data122998/Features_competition_train.tar.gz -C /home/aistudio/data/Features_train --strip-components 1 && rm -rf data122998/Features_competition_train.tar.gz
!tar -xf data123004/Features_competition_test_A.tar.gz -C /home/aistudio/data/Features_test --strip-components 1 && rm -rf data123004/Features_competition_test_A.tar.gz 
# 复制训练标签到data目录下%cp data122998/label_cls14_train.json /home/aistudio/data

%cd ~/work/# 从Github上下载PaddleVideo代码#!git clone https://github.com/PaddlePaddle/PaddleVideo.git# 若网速较慢，可使用如下方法下载!git clone https://hub.fastgit.org/PaddlePaddle/PaddleVideo.git

# 进入到PaddleVideo目录下%cd ~/work/PaddleVideo/

# 检查源代码文件结构!tree /home/aistudio/work/ -L 2

# 配置PaddleVideo环境!python3.7 -m pip install --upgrade pip
!python3.7 -m pip install --upgrade -r requirements.txt# 配置BMN前处理环境# %cd ~/work/BMN/ # !python3.7 -m pip install --upgrade -r requirements.txt

#数据格式
{
  "5679b8ad4eac486cbac82c4c496e278d_133.56_141.56": {     #视频名称_片段起始时间_片段结束时间(s)
          "duration_second": 8.0,
          "duration_frame": 200,
          "feature_frame": 200,
          "subset": "train",
          "annotations": [
              {
                  "segment": [
                      6.36,#动作起始时间
                      8.0  #动作结束时间
                  ],
                  "label": "11.0",
                  "label_name": "普通"
              }
          ]
      },
      ...
}

dataset = "/home/aistudio/data"feat_dir = dataset + '/Features_train'out_dir = dataset + '/Input_for_bmn'label_files = {    'train': 'label_cls14_train.json',    'validation': 'label_cls14_val.json'  }

# 数据预处理%cd /home/aistudio/work/PaddleVideo/applications/TableTennis/# 生成验证集!python3.7 val_split.py# 生成bmn训练数据和标签!python3.7 get_instance_for_bmn.py

# 存放地址检查 -- BMN训练输入数据和标签!tree /home/aistudio/data/Input_for_bmn/ -L 1

###
url = '/home/aistudio/data/Input_for_bmn/feature/'
###

###
with open('/home/aistudio/data/Input_for_bmn/label.json') as f:
    data = json.load(f)
###

###
jsonFile = open('/home/aistudio/data/Input_for_bmn/label_fixed.json', 'w')
###

# 运行标签修正脚本!python3.7 fix_bad_label.py

#4卡训练%cd ~/work/PaddleVideo/
export CUDA_VISIBLE_DEVICES=0,1,2,3python -B -m paddle.distributed.launch --gpus="0,1,2,3"  --log_dir=log_bmn main.py  --validate -c /home/aistudio/work/PaddleVideo/applications/TableTennis/configs/bmn_tabletennis.yaml

python -B main.py  --validate -c /home/aistudio/work/PaddleVideo/applications/TableTennis/configs/bmn_tabletennis.yaml

###DATASET:                                            #DATASET field
  batch_size: 4                                 #single card bacth size
  test_batch_size: 1
  num_workers: 8
  train:    format: "BMNDataset"
    file_path: "/home/aistudio/data/Input_for_bmn/label_fixed.json"
    subset: "train"######PIPELINE:                                           #PIPELINE field
  train:                                            #Mandotary, indicate the pipeline to deal with the training data
    load_feat:
      name: "LoadFeat"
      feat_path: "/home/aistudio/data/Input_for_bmn/feature"###

# 回到PaddleVideo主目录下%cd ~/work/PaddleVideo/ 
# 运行训练脚本!python3.7 -B main.py  --validate -c /home/aistudio/work/PaddleVideo/applications/TableTennis/configs/bmn_tabletennis.yaml# !python3.7 -B main.py  --validate -c /home/aistudio/work/BMN/configs/bmn_tabletennis_v2.0.yaml

# -c 指定配置文件，-p 指定训练好的模型参数， -o 指定模型输出目录 !python3.7 tools/export_model.py -c /home/aistudio/work/PaddleVideo/applications/TableTennis/configs/bmn_tabletennis.yaml -p /home/aistudio/work/PaddleVideo/output/BMN/BMN_epoch_00001.pdparams -o inference/BMN

# 跳转到脚本目录下%cd /home/aistudio/work/PaddleVideo/applications/TableTennis/extractor/

###sys.path.append(    "/home/aistudio/work/PaddleVideo/applications/TableTennis/predict/action_detect/")######
    dataset_dir = '/home/aistudio/data/Features_test/'
    output_dir = '/home/aistudio/data'###

###BMN:
    name: "BMN"
    model_file: "/home/aistudio/work/PaddleVideo/inference/BMN/BMN.pdmodel"
    params_file: "/home/aistudio/work/PaddleVideo/inference/BMN/BMN.pdiparams"###

# 运行推理预测脚本!python3.7 extract_bmn_for_tabletennis.py

# 跳转至脚本目录下%cd /home/aistudio/work/PaddleVideo/applications/TableTennis/datasets/script/

###with open('/home/aistudio/data/Output_for_bmn/prop.json') as f:
    data = json.load(f)######jsonFile = open('/home/aistudio/data/Output_for_bmn/submission.json', 'w')###

# 运行脚本!python3.7 submission_format_transfer.py