Hooks registry

Training lifecycle hooks. See Add a hook.

Build any of these from a config with hooks=[dict(type=...)], using the type string in the first column. Generated from the live registry - 26 classes (26 names including aliases), grouped by role.

Checkpointing

Save and load checkpoints / resume state (delegate to CheckpointManager).

Name	Summary
CheckpointLoader	Load model weights and, when requested, resume optimizer/train state.
CheckpointSaver	Save epoch/metric-oriented checkpoints during and after training.
CheckpointSaverIteration	Save iteration-oriented checkpoints on a pure global-step cadence.

pimm.engines.hooks.checkpoint.CheckpointLoader	Load model weights and, when requested, resume optimizer/train state.
pimm.engines.hooks.checkpoint.CheckpointSaver	Save epoch/metric-oriented checkpoints during and after training.
pimm.engines.hooks.checkpoint.CheckpointSaverIteration	Save iteration-oriented checkpoints on a pure global-step cadence.

Logging

Per-step / per-epoch scalar logging and run naming.

Name	Summary
InformationWriter	Assemble per-step console logs and write scalar train metrics.
IterationTimer	Measure data/batch latency and append timing and ETA to iteration logs.
WandbNamer	Auto-generate cfg.wandb_run_name from selected config values.

pimm.engines.hooks.logging.InformationWriter	Assemble per-step console logs and write scalar train metrics.
pimm.engines.hooks.logging.IterationTimer	Measure data/batch latency and append timing and ETA to iteration logs.
pimm.engines.hooks.logging.WandbNamer	Auto-generate cfg.wandb_run_name from selected config values.

Evaluation

In-loop validators, final testers, and SSL probes; they set the checkpoint-selection metric.

Name	Summary
FinalEvaluator	Run the configured tester once after training, usually on model_best.
InstanceSegmentationEvaluator	Instance-level segmentation metrics including ARI and detection/class stats.
MAEEvaluator	Validation hook for masked-autoencoder (MAE) pretraining.
PretrainEvaluator	Evaluate frozen pretraining features with downstream linear probes.
SemSegEvaluator	Evaluate point-wise semantic segmentation on the validation loader.

pimm.engines.hooks.eval.final_tester.FinalEvaluator	Run the configured tester once after training, usually on model_best.
pimm.engines.hooks.eval.instance_segmentation.InstanceSegmentationEvaluator	Instance-level segmentation metrics including ARI and detection/class stats.
pimm.engines.hooks.eval.pretrain.mae.MAEEvaluator	Validation hook for masked-autoencoder (MAE) pretraining.
pimm.engines.hooks.eval.pretrain.semantic_segmentation_pretrain.PretrainEvaluator	Evaluate frozen pretraining features with downstream linear probes.
pimm.engines.hooks.eval.semantic_segmentation.SemSegEvaluator	Evaluate point-wise semantic segmentation on the validation loader.

Diagnostics

Gradient/feature/prototype/parameter monitors and dtype/anneal controls.

Name	Summary
AttentionMaskAnnealingHook	Drive and log attention-mask annealing for the Panda detector decoder.
DtypeOverrider	Force matched layers to compute (and optionally store params) in a dtype.
FeatureStdMonitor	Monitor feature standard deviation to detect representation collapse.
GradientNormLogger	Log model gradient norms to the writer after each training step.
LogitEntropyLogger	Log the entropy of teacher logits using Sonata’s temperature schedule.
ParameterCounter	Log a parameter-count breakdown of the model at the start of training.
PrototypeUsageLogger	Monitor prototype utilization in Sonata-style clustering heads.
SimulateCrash	Hard-exit the process after a fixed number of steps to simulate a crash.

pimm.engines.hooks.diagnostics.AttentionMaskAnnealingHook	Drive and log attention-mask annealing for the Panda detector decoder.
pimm.engines.hooks.diagnostics.DtypeOverrider	Force matched layers to compute (and optionally store params) in a dtype.
pimm.engines.hooks.diagnostics.FeatureStdMonitor	Monitor feature standard deviation to detect representation collapse.
pimm.engines.hooks.diagnostics.GradientNormLogger	Log model gradient norms to the writer after each training step.
pimm.engines.hooks.diagnostics.LogitEntropyLogger	Log the entropy of teacher logits using Sonata's temperature schedule.
pimm.engines.hooks.diagnostics.ParameterCounter	Log a parameter-count breakdown of the model at the start of training.
pimm.engines.hooks.diagnostics.PrototypeUsageLogger	Monitor prototype utilization in Sonata-style clustering heads.
pimm.engines.hooks.diagnostics.SimulateCrash	Hard-exit the process after a fixed number of steps to simulate a crash.

Optimizer schedules

Hooks that mutate optimizer parameter groups (weight-decay exclusion / scheduling).

Name	Summary
WeightDecayExclusion	Rewrite optimizer param groups to exclude selected params from weight decay.
WeightDecayScheduler	Apply a cosine schedule to optimizer parameter-group weight decay.

pimm.engines.hooks.optimizer.WeightDecayExclusion	Rewrite optimizer param groups to exclude selected params from weight decay.
pimm.engines.hooks.optimizer.WeightDecayScheduler	Apply a cosine schedule to optimizer parameter-group weight decay.

Hugging Face export

Push checkpoints to the Hub during or after training.

Name	Summary
PushToHub	Push model checkpoints to the Hugging Face Hub during and/or after training.

pimm.engines.hooks.export.PushToHub

Push model checkpoints to the Hugging Face Hub during and/or after training.

Resources & profiling

Memory/GC management, resource-utilization logging, and the runtime profiler.

Name	Summary
GarbageHandler	Control Python garbage collection (and CUDA cache) on a step cadence.
ResourceUtilizationLogger	Log GPU and CPU/RAM utilization to the writer over the course of training.
RuntimeProfiler	Run a short torch.profiler trace before normal training proceeds.

pimm.engines.hooks.resources.GarbageHandler	Control Python garbage collection (and CUDA cache) on a step cadence.
pimm.engines.hooks.resources.ResourceUtilizationLogger	Log GPU and CPU/RAM utilization to the writer over the course of training.
pimm.engines.hooks.profiling.RuntimeProfiler	Run a short torch.profiler trace before normal training proceeds.

Core

Lifecycle plumbing shared by other hooks.

Name	Summary
ModelHook	Bridge that forwards lifecycle calls to a model implementing HookBase.

pimm.engines.hooks.default.ModelHook

Bridge that forwards lifecycle calls to a model implementing HookBase.