Pipeline Builder API

The Pipeline Builder API is the Pythonic interface for building ML inference pipelines on the Axelera Metis AIPU. It builds a pipeline from composable Python operators, in place of a YAML pipeline configuration.

The pages below are generated from the source docstrings.

Note:

The Operator classes each have a snake_case alias. e.g.:

• op.Seq can also be used as op.seq
• op.CenterCrop can also be used as op.center_crop
• etc.

Getting started​

• Quickstart: overview and quick start
• Model Compilation: compile your model to .axm
• Model Formats: .axm and .axe, what each holds and when to use which
• Pipeline Overview: build pipelines with full examples for every task type
• Cascade Tutorial: run a second model on each detection of the first
• Coordinate System Tutorial: where a box's coordinates live, and how a cascade composes them

Operator Reference​

• Transforms: image preprocessing (resize, letterbox, normalize, color convert)
• Image Preprocessing: geometric correction (op.perspective, op.barrel_correct)
• Postprocess: decode raw model output, NMS, coordinate transform, filter, top-k, softmax
• Results: convert arrays to typed objects (DetectedObject, PoseObject, etc.)
• Inference: model loading (op.load, op.onnx_model, op.pt2_model)
• AXE packages: load a saved .axe pipeline package (op.load_axe, AxeOperator, IncompatibleAxeError)
• Tracker: multi-object tracking (ByteTrack, OC-SORT, SORT, TrackTrack)
• Combinators: pipeline composition (op.seq, op.par, op.for_each, op.pack, op.unpack, op.itemgetter)

Model and Device Settings​

• Postamble: name a kernel for the model's postamble, in place of the onnxruntime tail (see also the Postamble guide)
• Device config: board settings a model needs before it runs, such as the 110m MVM limiter

Running Pipelines​

• Sources: video, camera and image input (cv.create_source, VideoSource, Image)
• Scheduler: device selection and execution behind pipeline.stream and pipeline.batch

Measuring​

• Benchmark: latency, throughput and device throughput (bench.benchmark_pipeline, behind a recipe's --benchmark)
• Profiler: per-operator and per-stage timing behind that report (OperatorProfiler, ResourceMonitor)
• Profiling: how to read the report those fields produce

Type Reference​

• Types: data types such as BBox, DetectedObject, PoseObject, Keypoint, SegmentedObject, TrackedObject, Classification