inference

sleap_nn.inference

Inference-related modules.

Quick start::

from sleap_nn.inference import predict, Predictor

# One-liner: source + model paths → Labels
labels = predict("video.mp4", model_paths=["/path/to/model"])

# Two-step: build once, predict many times with different settings
predictor = Predictor.from_model_paths(["/path/to/model"], device="cuda")
labels = predictor.predict("video.mp4", peak_threshold=0.3)

Modules:

Name	Description
bottomup	Inference modules for BottomUp models.
centroid_convert	Conversion helpers for centroid-only output representation.
embedding	Inference for the embedding model type: crops -> appearance vectors.
filters	Post-inference filtering: FilterConfig + FilterPipeline.
identity	Backward-compatibility re-export shim for identity ops.
layers	Inference layers — model-type-aware wrappers around a runtime backend.
loaders	Standalone checkpoint loading for the inference pipeline.
ops	Pure-ops library for inference (PR 1 of #508).
outputs	Outputs — the structured container produced by every InferenceLayer.
paf_grouping	Backward-compatibility re-export shim for PAF grouping.
peak_finding	Backward-compatibility re-export shim for peak finding.
postprocessing	Backward-compatibility re-export shim for post-inference filters.
predictor	Predictor — high-level orchestrator for the inference stack.
predictors	Predictors for running inference.
preprocess_info	PreprocInfo — preprocessing metadata used to reverse coordinate transforms.
provenance	Provenance metadata utilities for inference outputs.
providers	Provider protocol + concrete data sources for Predictor.
run	Top-level predict — one-call inference from model paths to Labels.
sam	SAM-powered prompted instance segmentation for INFERENCE.
segmentation	Inference utilities for bottom-up instance segmentation.
segmentation_convert	sleap-nn <-> sleap-io conversion helpers for instance segmentation masks.
single_instance	Inference modules for SingleInstance models.
streaming	Streaming primitives for the new inference stack.
tile_merger	Torch-native tiled inference: importance windows + accumulate/normalize canvas.
topdown	Inference modules for TopDown centroid and centered-instance models.
tracking	Tracking integration for the new Predictor flow.
utils	Miscellaneous utility functions for Inference modules.
writer	IncrementalLabelsWriter — buffered write of a .slp.