mlreco.utils.ppn module

mlreco.utils.ppn.contains(meta, point, point_type='3d')[source]

Decides whether a point is contained in the box defined by meta.

Parameters

  • meta (larcv::Voxel3DMeta or larcv::ImageMeta) –

  • point (larcv::Point3D or larcv::Point2D) –

  • point_type (str, optional) – Has to be “3d” for 3D, otherwise anything else works for 2D.

Returns

Return type

bool

mlreco.utils.ppn.pass_particle(gt_type, start, end, energy_deposit, vox_count)[source]

Filters particles based on their type, voxel count and energy deposit.

Parameters

  • gt_type (int) –

  • start (larcv::Point3D) –

  • end (larcv::Point3D) –

  • energy_deposit (float) –

  • vox_count (int) –

Returns

Return type

bool

Notes

Made during DUNE Pi0 workshop (?), do we need to keep it here? Assumes 3D

mlreco.utils.ppn.get_ppn_info(particle_v, meta, point_type='3d', min_voxel_count=5, min_energy_deposit=0, use_particle_shape=True)[source]

Gets particle points coordinates and informations for running PPN.

Parameters

  • particle_v –

  • meta (larcv::Voxel3DMeta or larcv::ImageMeta) –

  • point_type (str, optional) –

  • min_voxel_count (int, optional) –

  • min_energy_deposit (float, optional) –

Returns

Array of points of shape (N, 11) where 11 = x,y,z + point type + pdg code + energy deposit + num voxels + energy_init + particle index + start (0) or end (1) point tagging

Return type

np.array

Notes

We skip some particles under specific conditions (e.g. low energy deposit, low voxel count, nucleus track, etc.) For now in 2D we assume a specific 2d projection (plane).

mlreco.utils.ppn.nms_numpy(im_proposals, im_scores, threshold, size)[source]

Runs NMS algorithm on a list of predicted points and scores.

Parameters

  • im_proposals (np.array) – Shape (N, data_dim). Predicted points.

  • im_scores (np.array) – Shape (N, 2). Predicted scores.

  • threshold (float) – Threshold for overlap

  • size (int) – Half side of square window defined around each point

Returns

boolean array of same length as points/scores

Return type

np.array

mlreco.utils.ppn.group_points(ppn_pts, batch, label)[source]

if there are multiple ppn points in a very similar location, return the average pos

Parameters

  • ppn_pts (np.array) –

  • batch (np.array) –

  • label (np.array) –

Returns

Return type

np.array

mlreco.utils.ppn.uresnet_ppn_type_point_selector(data, out, score_threshold=0.5, type_score_threshold=0.5, type_threshold=1.999, entry=0, score_pool='max', enforce_type=True, batch_col=0, coords_col=(1, 4), type_col=(3, 8), score_col=(8, 10), selection=None, num_classes=5, apply_deghosting=True, **kwargs)[source]

Postprocessing of PPN points.

Parameters

  • - 5-types sparse tensor (data) –

  • - output dictionary of the full chain (out) –

  • - minimal detection score (score_threshold) –

  • - minimal score for a point type prediction to be considered (type_score_threshold) –

  • - distance threshold for matching w/ semantic type prediction (type_threshold) –

  • - which index to look at (within a batch of events) (entry) –

  • - which operation to use to pool PPN points scores (max/min/mean) (score_pool) –

  • - whether to force PPN points predicted of type X (enforce_type) – to be within N voxels of a voxel with same predicted semantic

  • - list of list of indices to consider exclusively (eg to get PPN predictions (selection) – within a cluster). Shape Batch size x N_voxels (not square)

Returns

  • [bid,x,y,z,score softmax values (2 columns), occupancy,

  • type softmax scores (5 columns), predicted type,

  • (optional) endpoint type]

  • 1 row per ppn-predicted points

mlreco.utils.ppn.uresnet_ppn_point_selector(data, out, nms_score_threshold=0.8, entry=0, window_size=4, score_threshold=0.9, **kwargs)[source]

Basic selection of PPN points.

Parameters

  • - 5-types sparse tensor (data) –

  • - ppn output (out) –

Returns

Return type

[x,y,z,bid,label] of ppn-predicted points

mlreco.utils.ppn.get_track_endpoints_geo(data, f, points_tensor=None, use_numpy=False)[source]

Compute endpoints of a track-like cluster f based on PPN point predictions (coordinates and scores) and geometry (voxels farthest apart from each other in the cluster).

If points_tensor is left unspecified, the endpoints will be purely based on geometry.

Input: - data is the input data tensor, which can be indexed by f. - points_tensor is the output of PPN ‘points’ (optional) - f is a list of voxel indices for voxels that belong to the track.

Output: - array of shape (2, 3) (2 endpoints, 3 coordinates each)

mlreco.utils.numba module mlreco.utils.track_clustering module