mmpose记录-config

以configs\body\2d_kpt_sview_rgb_img\deeppose\coco\res50_coco_256x192_rle.py为列

• learning policy

等待补充~

 

• model setting

model的设置

# model settings
model = dict(
    type='TopDown', # 首先模型确定detector，对应任务的明确。此处表从上至下的2D检测方法，其他的有如"mesh"的3D姿态估计的方法
    pretrained='torchvision://resnet50', # 预训练选项，默认为none，自己重头训练
    backbone=dict(type='ResNet', depth=50, num_stages=4, out_indices=(3, )),# 由type定了以后需提供的参数，TopDown中需要给出backbone
    neck=dict(type='GlobalAveragePooling'), # neck部分，此处RLE中有一个GAP操作故有'GlobalAveragePooling'
    keypoint_head=dict(    # 检测头部分,后面的部分同上
        type='DeepposeRegressionHead',
        in_channels=2048,
        num_joints=channel_cfg['num_output_channels'],
        loss_keypoint=dict(    # 不同于backbone常用预训练参数，检测头一般要训练，所以要设置好参数
            type='RLELoss',
            use_target_weight=True,
            size_average=True,
            residual=True),
        out_sigma=True),
    train_cfg=dict(),
    test_cfg=dict(flip_test=True, regression_flip_shift=True))
...

以下部分都在mmpose\models 中的不同目录下可找到 

TopDown初始化

#TopDown
class TopDown(BasePose):
    """Top-down pose detectors.

    Args:
        backbone (dict): Backbone modules to extract feature.
        keypoint_head (dict): Keypoint head to process feature.
        train_cfg (dict): Config for training. Default: None.
        test_cfg (dict): Config for testing. Default: None.
        pretrained (str): Path to the pretrained models.
        loss_pose (None): Deprecated arguments. Please use
            `loss_keypoint` for heads instead.
    """
    def __init__(self,
                 backbone,
                 neck=None,
                 keypoint_head=None,
                 train_cfg=None,
                 test_cfg=None,
                 pretrained=None,
                 loss_pose=None):
        ...

DeepposeRegressionHead初始化

#DeepposeRegressionHead
class DeepposeRegressionHead(nn.Module):
    """Deeppose regression head with fully connected layers.

    "DeepPose: Human Pose Estimation via Deep Neural Networks".

    Args:
        in_channels (int): Number of input channels
        num_joints (int): Number of joints
        loss_keypoint (dict): Config for keypoint loss. Default: None.
        out_sigma (bool): Predict the sigma (the viriance of the joint
            location) together with the joint location. Default: False
    """
    def __init__(self,
                 in_channels,
                 num_joints,
                 loss_keypoint=None,
                 out_sigma=False,
                 train_cfg=None,
                 test_cfg=None):
        ...

RLE初始化

#RLE
class RLELoss(nn.Module):
    """RLE Loss.

    `Human Pose Regression With Residual Log-Likelihood Estimation
    arXiv: <https://arxiv.org/abs/2107.11291>`_.

    Code is modified from `the official implementation
    <https://github.com/Jeff-sjtu/res-loglikelihood-regression>`_.

    Args:
        use_target_weight (bool): Option to use weighted MSE loss.
            Different joint types may have different target weights.
        size_average (bool): Option to average the loss by the batch_size.
        residual (bool): Option to add L1 loss and let the flow
            learn the residual error distribution.
        q_dis (string): Option for the identity Q(error) distribution,
            Options: "laplace" or "gaussian"
    """
    def __init__(self,
                 use_target_weight=False,
                 size_average=True,
                 residual=True,
                 q_dis='laplace'):
        super(RLELoss, self).__init__()
        self.size_average = size_average
        self.use_target_weight = use_target_weight
        self.residual = residual
        self.q_dis = q_dis

        self.flow_model = RealNVP()