experimenting.models package

Submodules

experimenting.models.losses module

Losses implementations

class experimenting.models.losses.HeatmapLoss(reduction='mask_mean', n_joints=13)

from https://github.com/anibali/margipose

class experimenting.models.losses.PixelWiseLoss(reduction='mask_mean', divergence=True)

from https://github.com/anibali/margipose/

class experimenting.models.losses.MultiPixelWiseLoss(reduction='mask_mean', divergence=True)

from https://github.com/anibali/margipose

experimenting.models.margipose module

Implementation of Margipose Model. Thanks to A. Nibali — original source code: https://github.com/anibali/margipose/src/margipose

class experimenting.models.margipose.MargiPoseModel2D(n_stages, backbone_path, n_joints, n_channels=1)

class HeatmapCombiner(n_joints, n_planes, out_channels)

forward(inputs)

Model forward process :param inputs: input batch

Returns

Triple of list of heatmaps of length {n_stages}

class experimenting.models.margipose.MargiPoseModel3D(n_stages, in_cnn, in_shape, n_joints, permute_axis=False)

Multi-stage marginal heatmap estimator

class HeatmapCombiner(n_joints, n_planes, out_channels)

forward(inputs) → Tuple[List[Any], List[Any], List[Any]]

Model forward process :param inputs: input batch

Returns

Triple of list of heatmaps of length {n_stages}

class experimenting.models.margipose.MargiPoseStage(n_joints, mid_shape, heatmap_space, permute)

experimenting.models.metrics module

Metrics implementation for 3D human pose comparisons

class experimenting.models.metrics.MPJPE(reduction=None, confidence=0, **kwargs)

forward(y_pr, points_gt, gt_mask=None)

Base forward method for metric evaluation :param y_pr: 3D prediction of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param points_gt: 3D gt of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param gt_mask: boolean mask, tensor of shape (BATCH_SIZExN_JOINTS). :param Applied to results, if provided:

Returns

Metric as single value, if reduction is given, or as a tensor of values

class experimenting.models.metrics.AUC(reduction=None, auc_reduction=torch.mean, start_at=0, end_at=500, step=30, **kwargs)

Area Under the Curve for PCK metric, at different thresholds (from 0 to 800)

forward(y_pr, points_gt, gt_mask=None)

Base forward method for metric evaluation :param y_pr: 3D prediction of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param points_gt: 3D gt of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param gt_mask: boolean mask, tensor of shape (BATCH_SIZExN_JOINTS). :param Applied to results, if provided:

Returns

Metric as single value, if reduction is given, or as a tensor of values

class experimenting.models.metrics.PCK(reduction=None, threshold=150, **kwargs)

Percentage of correct keypoints according to a thresold value. Usually default threshold is 150mm

forward(y_pr, points_gt, gt_mask=None)

Base forward method for metric evaluation :param y_pr: 3D prediction of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param points_gt: 3D gt of joints, tensor of shape (BATCH_SIZExN_JOINTSx3) :param gt_mask: boolean mask, tensor of shape (BATCH_SIZExN_JOINTS). :param Applied to results, if provided:

Returns

Metric as single value, if reduction is given, or as a tensor of values