experimenting.utils package¶

Submodules¶

experimenting.utils.cv_helpers module¶

experimenting.utils.cv_helpers.decay_heatmap(heatmap, sigma2=10)¶

Args

heatmap :: WxH matrix to decay
sigma2 :: (Default value = 1)

Returns

Heatmap obtained by gaussian-blurring the input

experimenting.utils.cv_helpers.get_heatmaps_steps(xyz, p_mat, width, height)¶

Args

xyz :: xyz coordinates as 3XNUM_JOINTS wrt world coord system
p_mat :: projection matrix from world to image plane
width :: width of the resulting frame
height :: height of the resulting frame

Returns

xyz wrf image coord system, uv image points of skeleton’s joints, uv mask

experimenting.utils.cv_helpers.decompose_projection_matrix(P)¶

QR decomposition of world2imageplane projection matrix

Args

P :: Projection matrix word 2 image plane

Returns

M matrix, camera matrix

experimenting.utils.cv_helpers.reproject_xyz_onto_world_coord(xyz, M, invert_z_axis=True)¶

Parameters

M – World to camera projection matrix
xyz – Skeleton joints as NUM_JOINTSx3

Returns: Skeleton joints reprojected in world coord system with shape NUM_JOINTSx3

experimenting.utils.cv_helpers.project_xyz_onto_camera_coord(xyz, M, invert_z_axis=True)¶

Args

xyz :: xyz coordinates as NUM_JOINTSx3 wrt world coord
M :: word2cam EXTRINSIC matrix

Returns

xyz coordinates projected onto cam coordinates system

experimenting.utils.dsntnn module¶

Differentiable DSNT operations for use in PyTorch computation graphs. Author: A. Nibali Url: https://github.com/anibali/margipose License: Apache

experimenting.utils.dsntnn.js_reg_losses(heatmaps, mu_t, sigma_t)¶

Calculate Jensen-Shannon divergences between heatmaps and target Gaussians.

Parameters

heatmaps (torch.Tensor) – Heatmaps generated by the model
mu_t (torch.Tensor) – Centers of the target Gaussians (in normalized units)
sigma_t (float) – Standard deviation of the target Gaussians (in pixels)

Returns

Per-location JS divergences.

experimenting.utils.dsntnn.dsnt(heatmaps)¶

Differentiable spatial to numerical transform.

Parameters: heatmaps (torch.Tensor) – Spatial representation of locations
Returns: Numerical coordinates corresponding to the locations in the heatmaps.

experimenting.utils.dsntnn.average_loss(losses, mask=None)¶

Calculate the average of per-location losses.

Parameters

losses (Tensor) – Predictions (B x L)
mask (Tensor, optional) – Mask of points to include in the loss calculation (B x L), defaults to including everything

experimenting.utils.nn_helpers module¶

Integration toolbox for pytorch nn package

class experimenting.utils.nn_helpers.FlatSoftmax¶

experimenting.utils.skeleton_helpers module¶

Skeleton wrapper. It provides a toolbox for plotting, projection, normalization, and denormalization of skeletons joints

experimenting.utils.train_helpers module¶

Toolbox for helping training and evaluation of agents

experimenting.utils.train_helpers.fit(cfg) → pytorch_lightning.Trainer¶

Launch training for a given config. Confs file can be found at /src/confs

Parameters: cfg (omegaconf.DictConfig) – Config dictionary

experimenting.utils.train_helpers.get_training_params(cfg: omegaconf.DictConfig)¶

Parameters: cfg (DictConfig :) – hydra configuration (examples in conf/train)