• mmpose系列（三）：中的hrnet_w48+deeppose的方法

    目前研究人体姿态的方法，参考上一篇博客，研究下hrnet是否可以通过deeppose的方法进行改造。

    改造内容如下：

    整个代码都是根据hrnet w48 384*288 来更改：

    原始的hrnet w48：

    1. _base_ = [
    2.     '../../../../_base_/default_runtime.py',
    3.     '../../../../_base_/datasets/coco.py'
    4. ]
    5. evaluation = dict(interval=10, metric='mAP', save_best='AP')
    6.  
    7. optimizer = dict(
    8.     type='Adam',
    9.     lr=5e-4,
    10. )
    11. optimizer_config = dict(grad_clip=None)
    12. # learning policy
    13. lr_config = dict(
    14.     policy='step',
    15.     warmup='linear',
    16.     warmup_iters=500,
    17.     warmup_ratio=0.001,
    18.     step=[170, 200])
    19. total_epochs = 210
    20. channel_cfg = dict(
    21.     num_output_channels=17,
    22.     dataset_joints=17,
    23.     dataset_channel=[
    24.         [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16],
    25.     ],
    26.     inference_channel=[
    27.         0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
    28.     ])
    29.  
    30. # model settings
    31. model = dict(
    32.     type='TopDown',
    33.     pretrained='https://download.openmmlab.com/mmpose/'
    34.     'pretrain_models/hrnet_w48-8ef0771d.pth',
    35.     backbone=dict(
    36.         type='HRNet',
    37.         in_channels=3,
    38.         extra=dict(
    39.             stage1=dict(
    40.                 num_modules=1,
    41.                 num_branches=1,
    42.                 block='BOTTLENECK',
    43.                 num_blocks=(4, ),
    44.                 num_channels=(64, )),
    45.             stage2=dict(
    46.                 num_modules=1,
    47.                 num_branches=2,
    48.                 block='BASIC',
    49.                 num_blocks=(4, 4),
    50.                 num_channels=(48, 96)),
    51.             stage3=dict(
    52.                 num_modules=4,
    53.                 num_branches=3,
    54.                 block='BASIC',
    55.                 num_blocks=(4, 4, 4),
    56.                 num_channels=(48, 96, 192)),
    57.             stage4=dict(
    58.                 num_modules=3,
    59.                 num_branches=4,
    60.                 block='BASIC',
    61.                 num_blocks=(4, 4, 4, 4),
    62.                 num_channels=(48, 96, 192, 384))),
    63.     ),
    64.     keypoint_head=dict(
    65.         type='TopdownHeatmapSimpleHead',
    66.         in_channels=48,
    67.         out_channels=channel_cfg['num_output_channels'],
    68.         num_deconv_layers=0,
    69.         extra=dict(final_conv_kernel=1, ),
    70.         loss_keypoint=dict(type='JointsMSELoss', use_target_weight=True)),
    71.     train_cfg=dict(),
    72.     test_cfg=dict(
    73.         flip_test=True,
    74.         post_process='default',
    75.         shift_heatmap=True,
    76.         modulate_kernel=11))
    77.  
    78. data_cfg = dict(
    79.     image_size=[288, 384],
    80.     heatmap_size=[72, 96],
    81.     num_output_channels=channel_cfg['num_output_channels'],
    82.     num_joints=channel_cfg['dataset_joints'],
    83.     dataset_channel=channel_cfg['dataset_channel'],
    84.     inference_channel=channel_cfg['inference_channel'],
    85.     soft_nms=False,
    86.     nms_thr=1.0,
    87.     oks_thr=0.9,
    88.     vis_thr=0.2,
    89.     use_gt_bbox=False,
    90.     det_bbox_thr=0.0,
    91.     bbox_file='data/coco/person_detection_results/'
    92.     'COCO_val2017_detections_AP_H_56_person.json',
    93. )
    94.  
    95. train_pipeline = [
    96.     dict(type='LoadImageFromFile'),
    97.     dict(type='TopDownGetBboxCenterScale', padding=1.25),
    98.     dict(type='TopDownRandomShiftBboxCenter', shift_factor=0.16, prob=0.3),
    99.     dict(type='TopDownRandomFlip', flip_prob=0.5),
    100.     dict(
    101.         type='TopDownHalfBodyTransform',
    102.         num_joints_half_body=8,
    103.         prob_half_body=0.3),
    104.     dict(
    105.         type='TopDownGetRandomScaleRotation', rot_factor=40, scale_factor=0.5),
    106.     dict(type='TopDownAffine'),
    107.     dict(type='ToTensor'),
    108.     dict(
    109.         type='NormalizeTensor',
    110.         mean=[0.485, 0.456, 0.406],
    111.         std=[0.229, 0.224, 0.225]),
    112.     dict(type='TopDownGenerateTarget', sigma=3),
    113.     dict(
    114.         type='Collect',
    115.         keys=['img', 'target', 'target_weight'],
    116.         meta_keys=[
    117.             'image_file', 'joints_3d', 'joints_3d_visible', 'center', 'scale',
    118.             'rotation', 'bbox_score', 'flip_pairs'
    119.         ]),
    120. ]
    121.  
    122. val_pipeline = [
    123.     dict(type='LoadImageFromFile'),
    124.     dict(type='TopDownGetBboxCenterScale', padding=1.25),
    125.     dict(type='TopDownAffine'),
    126.     dict(type='ToTensor'),
    127.     dict(
    128.         type='NormalizeTensor',
    129.         mean=[0.485, 0.456, 0.406],
    130.         std=[0.229, 0.224, 0.225]),
    131.     dict(
    132.         type='Collect',
    133.         keys=['img'],
    134.         meta_keys=[
    135.             'image_file', 'center', 'scale', 'rotation', 'bbox_score',
    136.             'flip_pairs'
    137.         ]),
    138. ]
    139.  
    140. test_pipeline = val_pipeline
    141.  
    142. data_root = 'data/coco'
    143. data = dict(
    144.     samples_per_gpu=32,
    145.     workers_per_gpu=2,
    146.     val_dataloader=dict(samples_per_gpu=32),
    147.     test_dataloader=dict(samples_per_gpu=32),
    148.     train=dict(
    149.         type='TopDownCocoDataset',
    150.         ann_file=f'{data_root}/annotations/person_keypoints_train2017.json',
    151.         img_prefix=f'{data_root}/train2017/',
    152.         data_cfg=data_cfg,
    153.         pipeline=train_pipeline,
    154.         dataset_info={{_base_.dataset_info}}),
    155.     val=dict(
    156.         type='TopDownCocoDataset',
    157.         ann_file=f'{data_root}/annotations/person_keypoints_val2017.json',
    158.         img_prefix=f'{data_root}/val2017/',
    159.         data_cfg=data_cfg,
    160.         pipeline=val_pipeline,
    161.         dataset_info={{_base_.dataset_info}}),
    162.     test=dict(
    163.         type='TopDownCocoDataset',
    164.         ann_file=f'{data_root}/annotations/person_keypoints_val2017.json',
    165.         img_prefix=f'{data_root}/val2017/',
    166.         data_cfg=data_cfg,
    167.         pipeline=test_pipeline,
    168.         dataset_info={{_base_.dataset_info}}),
    169. )

    主要改动有几个地方：

    model，pretrained可以设置为none,我自己就是设置的预训练模型为none进行训练的。

    必须加pooling层。

    并且修改对应的keypoints_head参数

    1. model = dict(
    2.     type='TopDown',
    3.     pretrained=r"E:/workspace/mmpose-master/pretrain_model/hrnet_w48-8ef0771d.pth",
    4.     backbone=dict(
    5.         type='HRNet',
    6.         in_channels=3,
    7.         extra=dict(
    8.             stage1=dict(
    9.                 num_modules=1,
    10.                 num_branches=1,
    11.                 block='BOTTLENECK',
    12.                 num_blocks=(4, ),
    13.                 num_channels=(64, )),
    14.             stage2=dict(
    15.                 num_modules=1,
    16.                 num_branches=2,
    17.                 block='BASIC',
    18.                 num_blocks=(4, 4),
    19.                 num_channels=(48, 96)),
    20.             stage3=dict(
    21.                 num_modules=4,
    22.                 num_branches=3,
    23.                 block='BASIC',
    24.                 num_blocks=(4, 4, 4),
    25.                 num_channels=(48, 96, 192)),
    26.             stage4=dict(
    27.                 num_modules=3,
    28.                 num_branches=4,
    29.                 block='BASIC',
    30.                 num_blocks=(4, 4, 4, 4),
    31.                 num_channels=(48, 96, 192, 384))),
    32.     ),
    33.     neck=dict(type='GlobalAveragePooling'),
    34.     keypoint_head=dict(
    35.         type='DeepposeRegressionHead',
    36.         in_channels=48,
    37.         num_joints=channel_cfg['num_output_channels'],
    38.         # num_deconv_layers=0,
    39.         # extra=dict(final_conv_kernel=1, ),
    40.         loss_keypoint=dict(type='SmoothL1Loss', use_target_weight=True)),
    41.     train_cfg=dict(),
    42.     test_cfg=dict(
    43.         flip_test=True,
    44.         post_process='default',
    45.         shift_heatmap=True,
    46.         modulate_kernel=11))

    train_pipeline: 将type='TopDownGenerateTarget', sigma=3  改为type='TopDownGenerateTargetRegression'即可

    1. train_pipeline = [
    2.     dict(type='LoadImageFromFile'),
    3.     dict(type='TopDownGetBboxCenterScale', padding=1.25),
    4.     dict(type='TopDownRandomShiftBboxCenter', shift_factor=0.16, prob=0.3),
    5.     dict(type='TopDownRandomFlip', flip_prob=0.5),
    6.     dict(
    7.         type='TopDownHalfBodyTransform',
    8.         num_joints_half_body=8,
    9.         prob_half_body=0.3),
    10.     dict(
    11.         type='TopDownGetRandomScaleRotation', rot_factor=40, scale_factor=0.5),
    12.     dict(type='TopDownAffine'),
    13.     dict(type='ToTensor'),
    14.     dict(
    15.         type='NormalizeTensor',
    16.         mean=[0.485, 0.456, 0.406],
    17.         std=[0.229, 0.224, 0.225]),
    18.     dict(type='TopDownGenerateTargetRegression'),
    19.     dict(
    20.         type='Collect',
    21.         keys=['img', 'target', 'target_weight'],
    22.         meta_keys=[
    23.             'image_file', 'joints_3d', 'joints_3d_visible', 'center', 'scale',
    24.             'rotation', 'bbox_score', 'flip_pairs'
    25.         ]),
    26. ]

    可以通过上篇文章生成onnx去验证模型。

    训练：

    单卡：

    python tools/train.py  configs/body25/deeppose/hrnet_w48_coco_384x288.py

    多卡：我是在服务器上用shell脚本训练的

    ./tools/dist_train.sh configs/body25/deeppose/hrnet_w48_coco_384x288.py  6

    这个是训练了10个迭代的效果。 正在训练后续给出训练好模型的mAP

    下一篇，给出hrnet+deeppose训练23点关键点和hrnet训练23点关键点的AP对比

  • 相关阅读:
    【车载开发系列】GIT教程---如何下载代码库
    Linux查看日志的几种命令（tail、cat、grep、head、sed...）
    使用Pytorch构建神经网络
    小礼盒礼金卡接口分享
    【python笔记】第十三节 常用模块
    单片机IAP固件升级分几步？（Qt上位机）
    NPDP产品经理国际认证报名有什么要求？
    25、Camunda Service Task、任务监听器和执行监听器
    Datawhale 2024 年 AI 夏令营第二期——基于术语词典干预的机器翻译挑战赛
    Python + Excel 办公自动化 01 —— 硬菜马上就来
  • 原文地址：https://blog.csdn.net/qq_38284951/article/details/126345138