fluttorch_gen 1.0.0

Generates a typed Dart API from a Fluttorch model manifest, so shapes, dtypes and preprocessing cannot drift from the training pipeline.

fluttorch_gen

Turns the manifest an export emitted into a Dart API, so that no shape, dtype or normalization constant is ever restated by hand on the device side.

Preprocessing written once in Python for training and again in Dart for serving drifts apart at the first refactor, and nothing tells you when it does. Here it is generated from the same document the model was exported with.

dev_dependencies:
  build_runner: ^2.4.0
  fluttorch_gen: ^0.3.0

Put two_layer.fluttorch.json next to your Dart sources and run the builder:

dart run build_runner build --delete-conflicting-outputs

What comes back is a class per model and a type per tensor:

final model = await TwoLayer.load(runtime, artifact: bytes);
final out = await model.run(features: TwoLayerFeatures(values));
print(out.score.values);

model.run(features: TwoLayerScore(...)) does not compile, and neither does passing a bare Tensor. The wrappers are extension types, so none of that costs anything at run time. A constructor refuses a list of the wrong length where the export fixed the dimension; where a dimension is dynamic there is withShape instead, deliberately wordier, because its shape is the one thing the compiler cannot check for you.

Two recorded preprocessing steps are refused rather than generated. resize and center_crop need to know which axes are spatial, the manifest records no tensor layout, and NCHW and NHWC would each produce a plausible and different answer. That is a gap in the schema rather than in the generator, and guessing there would reintroduce exactly the skew this package removes.

See the repository README for the whole pipeline.

License

Apache-2.0.