mlreco.models.layers.common.ppnplus module

class mlreco.models.layers.common.ppnplus.AttentionMask(score_threshold=0.5)[source]

Bases: torch.nn.modules.module.Module

Returns a masked tensor of x according to mask, where the number of coordinates between x and mask differ

__init__(score_threshold=0.5)[source]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x, mask)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

__module__ = 'mlreco.models.layers.common.ppnplus'
training: bool
class mlreco.models.layers.common.ppnplus.MergeConcat[source]

Bases: torch.nn.modules.module.Module

__init__()[source]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(input, other)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

__module__ = 'mlreco.models.layers.common.ppnplus'
training: bool
class mlreco.models.layers.common.ppnplus.ExpandAs[source]

Bases: torch.nn.modules.module.Module

__init__()[source]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x, shape, labels=None)[source]

x: feature tensor of input sparse tensor (N x F) labels: N x 0 tensor of labels

__module__ = 'mlreco.models.layers.common.ppnplus'
training: bool
class mlreco.models.layers.common.ppnplus.PPN(cfg, name='ppn')[source]

Bases: torch.nn.modules.module.Module

Point Proposal Network (PPN) implementation using MinkowskiEngine

It requires a UResNet network as a backbone.

Configuration

  • data_dim (int, default 3)

  • num_input (int, default 1)

  • allow_bias (bool, default False)

  • spatial_size (int, default 512)

  • leakiness (float, default 0.33)

  • activation (dict) – For activation function, defaults to {‘name’: ‘lrelu’, ‘args’: {}}

  • norm_layer (dict) – For normalization function, defaults to {‘name’: ‘batch_norm’, ‘args’: {}}

  • depth (int, default 5) – Depth of UResNet, also corresponds to how many times we down/upsample.

  • filters (int, default 16) – Number of filters in the first convolution of UResNet. Will increase linearly with depth.

  • reps (int, default 2) – Convolution block repetition factor

  • input_kernel (int, default 3) – Receptive field size for very first convolution after input layer.

  • num_classes (int, default 5)

  • score_threshold (float, default 0.5)

  • classify_endpoints (bool, default False) – Enable classification of points into start vs end points.

  • ppn_resolution (float, default 1.0)

  • ghost (bool, default False)

  • downsample_ghost (bool, default True)

  • use_true_ghost_mask (bool, default False)

  • mask_loss_name (str, default ‘BCE’) – Can be ‘BCE’ or ‘LogDice’

  • particles_label_seg_col (int, default -2) – Which column corresponds to particles’ semantic label

  • track_label (int, default 1)

Output

  • points (torch.Tensor) – Contains X, Y, Z predictions, semantic class prediction logits, and prob score

  • mask_ppn (list of torch.Tensor) – Binary mask at various spatial scales of PPN predictions (voxel-wise score > some threshold)

  • ppn_coords (list of torch.Tensor) – List of XYZ coordinates at various spatial scales.

  • ppn_layers (list of torch.Tensor) – List of score features at various spatial scales.

  • ppn_output_coordinates (torch.Tensor) – XYZ coordinates tensor at the very last layer of PPN (initial spatial scale)

  • classify_endpoints (torch.Tensor) – Logits for end/start point classification.

See also

PPNLonelyLoss, mlreco.models.uresnet_ppn_chain

__init__(cfg, name='ppn')[source]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(final, decoderTensors, ghost=None, ghost_labels=None)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

__module__ = 'mlreco.models.layers.common.ppnplus'
training: bool
class mlreco.models.layers.common.ppnplus.PPNLonelyLoss(cfg, name='ppn')[source]

Bases: torch.nn.modules.loss._Loss

Loss function for PPN.

Output

  • reg_loss (float) – Distance loss

  • mask_loss (float) – Binary voxel-wise prediction (is there an object of interest or not)

  • type_loss (float) – Semantic prediction loss.

  • classify_endpoints_loss (float)

  • classify_endpoints_acc (float)

See also

PPN, mlreco.models.uresnet_ppn_chain

__init__(cfg, name='ppn')[source]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

__module__ = 'mlreco.models.layers.common.ppnplus'
reduction: str
forward(result, segment_label, particles_label)[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

mlreco.models.layers.common.normalizations module mlreco.models.layers.common.senet module