Python API

Each component in a config is a dict whose type is a name registered here; pimm builds the class with the dict’s other keys. Summaries are the first paragraph of each class’s docstring. To add a name, see Add a component.

Registry	Names	Used for
Models	54	model.type and nested backbone.type
Datasets	7	data.<split>.type
Transforms	46	entries of transform lists
Losses	14	entries of criteria lists
Hooks	26	entries of hooks
Trainers	5	train.type
Testers	2	test.type
Optimizers	3	optimizer.type
Schedulers	6	scheduler.type

Models

Name	Class	Summary
DefaultClassifier	pimm.models.default.DefaultClassifier	—
DefaultInsSegmentor	pimm.models.default.DefaultInsSegmentor	—
DefaultSegmentor	pimm.models.default.DefaultSegmentor	—
DefaultSegmentorV2	pimm.models.default.DefaultSegmentorV2	—
DefaultSegmentorV3	pimm.models.default.DefaultSegmentorV3	—
detector-v3m2	pimm.models.panda_detector.detector_v3m2.MultiLabelDetectorV3M2	Shared-backbone Panda detector with one batched query decoder for all labels.
detector-v4	pimm.models.panda_detector.detector_v4.UnifiedDetector	Unified Panda detector: arbitrary per-query heads, optional pre-attention point removal, configurable queries, and per-label overlap.
detector-v5	pimm.models.panda_detector.detector_v5.UnifiedDetector	Backbone, shared query decoder, losses, and panoptic postprocessing.
detector-v5m2	pimm.models.panda_detector.detector_v5m2.UnifiedDetector	Backbone, shared query decoder, losses, and panoptic postprocessing.
DINOEnhancedSegmentor	pimm.models.default.DINOEnhancedSegmentor	—
LitePT	pimm.models.litept.litept.LitePT	—
LoRAAdapter	pimm.models.lora.LoRAAdapter	Freeze a built model and LoRA-tune its attention blocks.
MinkUNet101	pimm.models.sparse_unet.mink_unet.MinkUNet101	—
MinkUNet14	pimm.models.sparse_unet.mink_unet.MinkUNet14	—
MinkUNet14A	pimm.models.sparse_unet.mink_unet.MinkUNet14A	—
MinkUNet14B	pimm.models.sparse_unet.mink_unet.MinkUNet14B	—
MinkUNet14C	pimm.models.sparse_unet.mink_unet.MinkUNet14C	—
MinkUNet14D	pimm.models.sparse_unet.mink_unet.MinkUNet14D	—
MinkUNet18	pimm.models.sparse_unet.mink_unet.MinkUNet18	—
MinkUNet18A	pimm.models.sparse_unet.mink_unet.MinkUNet18A	—
MinkUNet18B	pimm.models.sparse_unet.mink_unet.MinkUNet18B	—
MinkUNet18D	pimm.models.sparse_unet.mink_unet.MinkUNet18D	—
MinkUNet34	pimm.models.sparse_unet.mink_unet.MinkUNet34	—
MinkUNet34A	pimm.models.sparse_unet.mink_unet.MinkUNet34A	—
MinkUNet34B	pimm.models.sparse_unet.mink_unet.MinkUNet34B	—
MinkUNet34C	pimm.models.sparse_unet.mink_unet.MinkUNet34C	—
MinkUNet50	pimm.models.sparse_unet.mink_unet.MinkUNet50	—
PG-v1m1	pimm.models.point_group.point_group_v1m1_base.PointGroup	—
PointTransformer-Cls26	pimm.models.point_transformer.point_transformer_cls.PointTransformerCls26	—
PointTransformer-Cls38	pimm.models.point_transformer.point_transformer_cls.PointTransformerCls38	—
PointTransformer-Cls50	pimm.models.point_transformer.point_transformer_cls.PointTransformerCls50	—
PointTransformer-PartSeg26	pimm.models.point_transformer.point_transformer_partseg.PointTransformerSeg26	—
PointTransformer-PartSeg38	pimm.models.point_transformer.point_transformer_partseg.PointTransformerSeg38	—
PointTransformer-PartSeg50	pimm.models.point_transformer.point_transformer_partseg.PointTransformerSeg50	—
PointTransformer-Seg26	pimm.models.point_transformer.point_transformer_seg.PointTransformerSeg26	—
PointTransformer-Seg38	pimm.models.point_transformer.point_transformer_seg.PointTransformerSeg38	—
PointTransformer-Seg50	pimm.models.point_transformer.point_transformer_seg.PointTransformerSeg50	—
PoLAr-MAE	pimm.models.polarmae.polarmae.PoLArMAE	Self-contained PoLAr-MAE for pimm’s DefaultTrainer.
PoLArMAE-SemSeg	pimm.models.polarmae.polarmae_semseg.PoLArMAESemSeg	PoLAr-MAE semantic segmentation for pimm’s DefaultTrainer.
PT-v2m1	pimm.models.point_transformer_v2.point_transformer_v2m1_origin.PointTransformerV2	—
PT-v2m2	pimm.models.point_transformer_v2.point_transformer_v2m2_base.PointTransformerV2	—
PT-v2m3	pimm.models.point_transformer_v2.point_transformer_v2m3_pdnorm.PointTransformerV2	—
PT-v3m1	pimm.models.point_transformer_v3.point_transformer_v3m1_base.PointTransformerV3	—
PT-v3m2	pimm.models.point_transformer_v3.point_transformer_v3m2_sonata.PointTransformerV3	—
PT-v3m4	pimm.models.point_transformer_v3.point_transformer_v3m8.PointTransformerV3	Point Transformer V3 with 3D RoPE and bottleneck CPE.
PT-v3m8	pimm.models.point_transformer_v3.point_transformer_v3m8.PointTransformerV3	Point Transformer V3 with 3D RoPE and bottleneck CPE.
Sonata-v1m1	pimm.models.sonata.sonata_v1m1_base.Sonata	—
Sonata-v1m2	pimm.models.sonata.sonata_v1m2_uni_teacher_head.Sonata	—
SpUNet-v1m1	pimm.models.sparse_unet.spconv_unet_v1m1_base.SpUNetBase	—
SpUNet-v1m2	pimm.models.sparse_unet.spconv_unet_v1m2_bn_momentum.SpUNetBase	—
SpUNet-v1m3	pimm.models.sparse_unet.spconv_unet_v1m3_pdnorm.SpUNetBase	—
SpUNetNoSkipBase	pimm.models.sparse_unet.spconv_unet_v1m1_base.SpUNetNoSkipBase	—
Volt-v1m1	pimm.models.volt.volt_v1m1.VoltBackbone	Flat Volt encoder for Sonata pretraining.
Volt-v1m2	pimm.models.volt.volt_v1m2.VoltBackboneV2	Flat Volt encoder + transposed-conv decoder, returning per-input-point features.

Datasets

Name	Class	Summary
ConcatDataset	pimm.datasets.defaults.ConcatDataset	Concatenate several configured datasets into a single index space.
DefaultDataset	pimm.datasets.defaults.DefaultDataset	Generic point-cloud dataset over preprocessed per-event .npy assets.
JAXTPCDataset	pimm.datasets.jaxtpc_dataset.JAXTPCDataset	Multimodal LArTPC simulation dataset over co-indexed JAXTPC HDF5 files.
LUCiDDataset	pimm.datasets.lucid_dataset.LUCiDDataset	Water Cherenkov detector dataset over co-indexed LUCiD HDF5 files.
PILArNetH5Dataset	pimm.datasets.pilarnet.h5.PILArNetH5Dataset	PILArNet-M LArTPC dataset read directly from clustered HDF5 shards.
PILArNetParquetDataset	pimm.datasets.pilarnet.parquet.PILArNetParquetDataset	Map-style PILArNet reader backed by parquet (Arrow-mmap).
PILArNetParquetIterableDataset	pimm.datasets.pilarnet.parquet.PILArNetParquetIterableDataset	Streaming PILArNet reader for large-scale pretraining.

Transforms

Name	Class	Summary
CenterShift	pimm.datasets.transform.spatial.CenterShift	Center coord on the midpoint of its bounding box per axis.
ChromaticAutoContrast	pimm.datasets.transform.color.ChromaticAutoContrast	Randomly auto-contrast the per-point RGB color.
ChromaticJitter	pimm.datasets.transform.color.ChromaticJitter	Add independent per-point Gaussian noise to RGB channels.
ChromaticTranslation	pimm.datasets.transform.color.ChromaticTranslation	Randomly shift all RGB channels by a shared offset.
ClipGaussianJitter	pimm.datasets.transform.spatial.ClipGaussianJitter	Add isotropic clipped multivariate-Gaussian noise to coord.
Collect	pimm.datasets.transform.base.Collect	Final projection of a sample dict into the model-facing batch contract.
ComputeAnchors	pimm.datasets.transform.instance.ComputeAnchors	Compute geometric anchors once per event and attach them to the sample.
ConditionalRandomTransform	pimm.datasets.transform.spatial.ConditionalRandomTransform	Wall-aware random translation that keeps points inside fixed bounds.
ContrastiveViewsGenerator	pimm.datasets.transform.multiview.ContrastiveViewsGenerator	Generate two independently-augmented views for contrastive SSL.
Copy	pimm.datasets.transform.base.Copy	Duplicate keys in the sample dict under new names.
CropBoundary	pimm.datasets.transform.spatial.CropBoundary	Drop points whose segment label is 0 or 1.
ElasticDistortion	pimm.datasets.transform.spatial.ElasticDistortion	Apply smooth random elastic warping to coord.
EnergeticTranslation	pimm.datasets.transform.color.EnergeticTranslation	Randomly shift per-point energy by a shared scalar offset.
EnergyJitter	pimm.datasets.transform.color.EnergyJitter	Apply multiplicative per-point jitter to energy.
GridSample	pimm.datasets.transform.spatial.GridSample	Voxel-downsample a point cloud onto a hash grid.
HardExampleCrop	pimm.datasets.transform.spatial.HardExampleCrop	Spherical crop biased toward rare (“hard”) segment labels.
HueSaturationTranslation	pimm.datasets.transform.color.HueSaturationTranslation	Randomly translate hue and scale saturation of point colors.
InstanceParser	pimm.datasets.transform.instance.InstanceParser	Compact instance ids and derive per-instance bounding boxes and centroids.
LocalCovarianceFeatures	pimm.datasets.transform.instance.LocalCovarianceFeatures	Per-point local-neighborhood covariance eigen-features.
LogTransform	pimm.datasets.transform.color.LogTransform	Compress scalar features (e.g. energy) onto [-1, 1].
MixedScaleGeometryMultiViewGenerator	pimm.datasets.transform.multiview.MixedScaleGeometryMultiViewGenerator	Multi-view generator with coarse locals plus extra fine local crops.
MomentumTransform	pimm.datasets.transform.color.MomentumTransform	Log10-compress strictly-positive momentum values.
MultiplicativeRandomJitter	pimm.datasets.transform.spatial.MultiplicativeRandomJitter	Multiply one or more keys by clipped Gaussian noise around 1.
MultiViewGenerator	pimm.datasets.transform.multiview.MultiViewGenerator	Generate multiple global and local crops for DINO/iBOT-style SSL.
NormalizeColor	pimm.datasets.transform.color.NormalizeColor	Scale per-point colors from [0, 255] into [0, 1].
NormalizeCoord	pimm.datasets.transform.spatial.NormalizeCoord	Recenter and rescale coord into a normalized frame.
PDGToSemantic	pimm.datasets.transform.detector.PDGToSemantic	Fallback semantic labels derived from PDG codes.
PointClip	pimm.datasets.transform.spatial.PointClip	Clamp coord to an axis-aligned point-cloud range.
PositiveShift	pimm.datasets.transform.spatial.PositiveShift	Translate coord so its minimum corner sits at the origin.
RandomColorDrop	pimm.datasets.transform.color.RandomColorDrop	Randomly drop (zero out or attenuate) per-point color.
RandomColorGrayScale	pimm.datasets.transform.color.RandomColorGrayScale	Randomly convert per-point RGB color to grayscale.
RandomColorJitter	pimm.datasets.transform.color.RandomColorJitter	Randomly jitter brightness, contrast, saturation, and hue of point colors.
RandomDrop	pimm.datasets.transform.spatial.RandomDrop	Randomly overwrite a fraction of one key’s rows with a constant.
RandomDropout	pimm.datasets.transform.spatial.RandomDropout	Randomly drop a fraction of points from the sample.
RandomFlip	pimm.datasets.transform.spatial.RandomFlip	Randomly mirror the cloud across one or more axes.
RandomJitter	pimm.datasets.transform.spatial.RandomJitter	Add clipped Gaussian noise to one or more point-aligned keys.
RandomRotate	pimm.datasets.transform.spatial.RandomRotate	Rotate the cloud by a random angle about one axis.
RandomRotateTargetAngle	pimm.datasets.transform.spatial.RandomRotateTargetAngle	Rotate the cloud by a random angle drawn from a discrete set.
RandomScale	pimm.datasets.transform.spatial.RandomScale	Scale coord by a random factor.
RandomShift	pimm.datasets.transform.spatial.RandomShift	Translate coord by a uniform random per-axis offset.
RelativeLogNormalize	pimm.datasets.transform.color.RelativeLogNormalize	Per-event relative log normalization (e.g. for hit times).
SetRandomValue	pimm.datasets.transform.spatial.SetRandomValue	Overwrite one or more keys with clipped Gaussian noise.
ShufflePoint	pimm.datasets.transform.spatial.ShufflePoint	Randomly permute the point ordering of the sample.
SphereCrop	pimm.datasets.transform.spatial.SphereCrop	Crop to the nearest points around a center to cap the point count.
ToTensor	pimm.datasets.transform.base.ToTensor	Recursively convert numpy arrays and numeric leaves to torch tensors.
Update	pimm.datasets.transform.base.Update	Inject constant key-value pairs into the sample dict.

Losses

Name	Class	Summary
BinaryFocalLoss	pimm.models.losses.misc.BinaryFocalLoss	Binary focal loss for class-imbalanced binary targets.
CrossEntropyHeadLoss	pimm.models.losses.misc.CrossEntropyHeadLoss	Cross-entropy for an extra categorical per-query head.
CrossEntropyLoss	pimm.models.losses.misc.CrossEntropyLoss	Standard multi-class cross-entropy over per-point class logits.
DiceLoss	pimm.models.losses.misc.DiceLoss	Soft multi-class Dice loss over per-point class probabilities.
FastInstanceSegmentationLoss	pimm.models.losses.instance_fast.FastInstanceSegmentationLoss	Mask2Former-style instance segmentation loss (cached, vectorized).
FastInstanceSegmentationRegressionLoss	pimm.models.losses.instance_regression_fast.FastInstanceSegmentationRegressionLoss	Instance segmentation loss plus configurable per-instance regressions.
FastUnifiedInstanceLoss	pimm.models.losses.instance_unified_fast.FastUnifiedInstanceLoss	Combine instance mask/class losses with configured query-head losses.
FocalLoss	pimm.models.losses.misc.FocalLoss	Multi-class focal loss over per-point class logits.
InstanceSegmentationLoss	pimm.models.losses.instance_fast.InstanceSegmentationLoss	DEPRECATED alias for :class:FastInstanceSegmentationLoss.
L1RegressionLoss	pimm.models.losses.misc.L1RegressionLoss	Mean-absolute-error (L1) regression loss for continuous targets.
LovaszLoss	pimm.models.losses.lovasz.LovaszLoss	Lovasz loss: a direct surrogate for the IoU (Jaccard) metric.
MSERegressionLoss	pimm.models.losses.misc.MSERegressionLoss	Mean-squared-error (L2) regression loss for continuous targets.
SmoothCELoss	pimm.models.losses.misc.SmoothCELoss	Label-smoothed cross-entropy with NaN-robust averaging.
SmoothL1RegressionLoss	pimm.models.losses.misc.SmoothL1RegressionLoss	Smooth-L1 (Huber) regression loss for continuous targets.

Hooks

Name	Class	Summary
AttentionMaskAnnealingHook	pimm.engines.hooks.diagnostics.AttentionMaskAnnealingHook	Drive and log attention-mask annealing for the Panda detector decoder.
CheckpointLoader	pimm.engines.hooks.checkpoint.CheckpointLoader	Load model weights and, when requested, resume optimizer/train state.
CheckpointSaver	pimm.engines.hooks.checkpoint.CheckpointSaver	Save epoch/metric-oriented checkpoints during and after training.
CheckpointSaverIteration	pimm.engines.hooks.checkpoint.CheckpointSaverIteration	Save iteration-oriented checkpoints on a pure global-step cadence.
DtypeOverrider	pimm.engines.hooks.diagnostics.DtypeOverrider	Force matched layers to compute (and optionally store params) in a dtype.
FeatureStdMonitor	pimm.engines.hooks.diagnostics.FeatureStdMonitor	Monitor feature standard deviation to detect representation collapse.
FinalEvaluator	pimm.engines.hooks.eval.final_tester.FinalEvaluator	Run the configured tester once after training, usually on model_best.
GarbageHandler	pimm.engines.hooks.resources.GarbageHandler	Control Python garbage collection (and CUDA cache) on a step cadence.
GradientNormLogger	pimm.engines.hooks.diagnostics.GradientNormLogger	Log model gradient norms to the writer after each training step.
InformationWriter	pimm.engines.hooks.logging.InformationWriter	Assemble per-step console logs and write scalar train metrics.
InstanceSegmentationEvaluator	pimm.engines.hooks.eval.instance_segmentation.InstanceSegmentationEvaluator	Instance-level segmentation metrics including ARI and detection/class stats.
IterationTimer	pimm.engines.hooks.logging.IterationTimer	Measure data/batch latency and append timing and ETA to iteration logs.
LogitEntropyLogger	pimm.engines.hooks.diagnostics.LogitEntropyLogger	Log the entropy of teacher logits using Sonata’s temperature schedule.
MAEEvaluator	pimm.engines.hooks.eval.pretrain.mae.MAEEvaluator	Validation hook for masked-autoencoder (MAE) pretraining.
ModelHook	pimm.engines.hooks.default.ModelHook	Bridge that forwards lifecycle calls to a model implementing HookBase.
ParameterCounter	pimm.engines.hooks.diagnostics.ParameterCounter	Log a parameter-count breakdown of the model at the start of training.
PretrainEvaluator	pimm.engines.hooks.eval.pretrain.semantic_segmentation_pretrain.PretrainEvaluator	Evaluate frozen pretraining features with downstream linear probes.
PrototypeUsageLogger	pimm.engines.hooks.diagnostics.PrototypeUsageLogger	Monitor prototype utilization in Sonata-style clustering heads.
PushToHub	pimm.engines.hooks.export.PushToHub	Push model checkpoints to the Hugging Face Hub during and/or after training.
ResourceUtilizationLogger	pimm.engines.hooks.resources.ResourceUtilizationLogger	Log GPU and CPU/RAM utilization to the writer over the course of training.
RuntimeProfiler	pimm.engines.hooks.profiling.RuntimeProfiler	Run a short torch.profiler trace before normal training proceeds.
SemSegEvaluator	pimm.engines.hooks.eval.semantic_segmentation.SemSegEvaluator	Evaluate point-wise semantic segmentation on the validation loader.
SimulateCrash	pimm.engines.hooks.diagnostics.SimulateCrash	Hard-exit the process after a fixed number of steps to simulate a crash.
WandbNamer	pimm.engines.hooks.logging.WandbNamer	Auto-generate cfg.wandb_run_name from selected config values.
WeightDecayExclusion	pimm.engines.hooks.optimizer.WeightDecayExclusion	Rewrite optimizer param groups to exclude selected params from weight decay.
WeightDecayScheduler	pimm.engines.hooks.optimizer.WeightDecayScheduler	Apply a cosine schedule to optimizer parameter-group weight decay.

Trainers

Name	Class	Summary
DefaultTrainer	pimm.engines.train.Trainer	Default single-dataset supervised/SSL trainer.
GRPOTrainer	pimm.engines.train.GRPOTrainer	Reinforcement-learning trainer implementing GRPO at the trainer level.
ImageClassTrainer	pimm.engines.train.ImageClassTrainer	Trainer for dense 2D image batches (e.g. rasterized ring images).
InsegTrainer	pimm.engines.train.InsegTrainer	Trainer for instance segmentation with instance-aware collation.
MultiDatasetTrainer	pimm.engines.train.MultiDatasetTrainer	Trainer that draws mixed batches from several datasets.

Testers

Name	Class	Summary
InstanceSegTester	pimm.engines.test.InstanceSegTester	Panoptic/Instance segmentation tester following InstanceSegmentationEvaluator pattern.
SemSegTester	pimm.engines.test.SemSegTester	Semantic segmentation tester following SemSegEvaluator pattern.

Optimizers

Name	Class	Summary
Adam	torch.optim.adam.Adam	Implements Adam algorithm.
AdamW	torch.optim.adamw.AdamW	Implements AdamW algorithm, where weight decay does not accumulate in the momentum nor variance.
SGD	torch.optim.sgd.SGD	Implements stochastic gradient descent (optionally with momentum).

Schedulers

Name	Class	Summary
CosineAnnealingLR	pimm.utils.scheduler.CosineAnnealingLR	Cosine annealing scheduler using total training steps as T_max.
ExpLR	pimm.utils.scheduler.ExpLR	Exponential decay scheduler normalized by total training steps.
MultiStepLR	pimm.utils.scheduler.MultiStepLR	Multi-step scheduler that accepts milestone ratios of total steps.
MultiStepWithWarmupLR	pimm.utils.scheduler.MultiStepWithWarmupLR	Multi-step decay with linear warmup expressed as total-step ratios.
OneCycleLR	pimm.utils.scheduler.OneCycleLR	torch.optim.lr_scheduler.OneCycleLR, Block total_steps
PolyLR	pimm.utils.scheduler.PolyLR	Polynomial decay scheduler over a fixed number of steps.