Model files

from BEAM import BeamModel

model.save("donor_A")                       # writes donor_A.beam
model = BeamModel.from_file("donor_A.beam")

A .beam file stores a model’s counts and its configuration (alphabet, order, maximum length, expert weights and prior strengths), so a loaded model gives exactly the same probabilities and can still be combined with other models. By default only observed contexts are stored. save adds .beam only when the path has no suffix; from_file reads any path and recognises the file by its contents.

Each file also records when the model was built (model.metadata()["created"]) and ends in a CRC-32C checksum of its contents: a damaged or truncated file is refused with ModelFileError rather than read wrongly. A sound file too large for the memory budget raises BudgetExceeded instead.

Save before compact(): a compacted model has no counts left to write (NotFinalized).

Building the same model twice writes the same file, apart from the build times and the final checksum. BEAM.file_provenance_span() returns the (offset, length) of the build times, (32, 24) in the current format, for byte-for-byte comparisons.

Compatibility

This version of BEAM reads and writes file format 9 (BEAM.pretrained.MODEL_FORMAT). The file format is part of the API: a release that writes a new format still reads the previous one, or it is a new major version. A file from a pre-release development build is refused with a message naming its format; rebuild it from its sequences.