Pn1F

Pn1F is an open, vendor-neutral file format and specification for describing neural network models and their execution graphs. Designed to enable portable inference across heterogeneous hardware, Pn1F aims to facilitate model exchange, reproducibility, and deployment in research and industry. The acronym Pn1F is the designation used in the specification; the exact expansion has not been publicly standardized.

Structure and features: A Pn1F file is a binary container consisting of a header, a graph section, a parameter block, and an optional metadata section. The header contains a magic string, a major/minor version, and a schema hash. The graph section encodes nodes, edges, and operation types using a compact intermediate representation. The parameter block stores tensors (weights and biases) with explicit shapes and data types. The metadata section can include runtime constraints, device preferences, and provenance information. The format supports compression, versioning, and a pluggable operator library to map abstract operations to backend implementations.

Development and usage: Pn1F was introduced in 2020 by the Pn1 Foundation as part of an initiative for standardized model interchange. It has since seen contributions from multiple research groups and hardware vendors. Toolchains exist for validating, converting to/from ONNX and other formats, and executing Pn1F models on various runtimes. The standard emphasizes deterministic behavior where possible and provides mechanisms for reproducibility, such as exact versioning and metadata provenance.

See also: ONNX, TensorFlow Lite, Core ML, Open Neural Network Exchange.