experimenting.agents package

Agents module

Agents package provides a toolbox agents for a multitude of tasks

Tasks provided

  • Autoencoder

  • Classification

  • Heatmap

  • 2D joint estimation

  • 3D joint estimation

Submodules

experimenting.agents.base module

Base class for agents classes. Each agent provides a dataset factory class to get train, val, and test datasets. Each agent must implement training, validation, and test steps methods as well as epoch_end methods

class experimenting.agents.base.BaseModule(hparams, dataset_constructor)
prepare_data()

It uses the provided factory constructor to get train, val, and test sets and set them as attributes

experimenting.agents.margipose_estimator module

class experimenting.agents.margipose_estimator.MargiposeEstimator(hparams, estimate_depth=False, test_metrics=None)

Agents for training and testing multi-stage 3d joints estimator using marginal heatmaps (denoted as Margipose)

denormalize_predictions(normalized_predictions, b_y)

b Denormalize skeleton prediction and reproject onto original coord system

Args:

normalized_predictions (torch.Tensor): normalized predictions b_y: batch y object (as returned by 3d joints dataset)

Returns:

Returns torch tensor of shape (BATCH_SIZE, NUM_JOINTS, 3)

Note:

Prediction skeletons are normalized according to batch depth value z_ref or torso length

Todo:

[] de-normalization is currently CPU only

experimenting.agents.margipose_estimator.predict3d(xy_hm, zy_hm, xz_hm)

Predict normalized 3d skeleton joints

Parameters

outs (list, list, list) – output of the model

Returns

torch tensor of normalized skeleton joints with shape (BATCH_SIZE, NUM_JOINTS, 3)

Note

prediction used dsnnt toolbox