KissDL

A deep learning model compiler based on K.I.S.S. principle, which takes in a trained model, and generates a cold hard, plain, self-contained C99 implementation.

Comparing to Tensorflow Lite, this compiler does two things that are missing in TF Lite:

• Memory optimization

  • Using Genetic Algorithm for optimal static memory allocation (see GAAlloc)

• C99 code generation

  • Parameters are pre-computed and injected into kernels' body

The trained model is saved as an emodel, which is just a collection of readable Erlang expressions, plus raw data files. emodel can be generated from TF Lite model.

Note: This is still under development.

Build

> cd(path of this compiler).
ok
> make:all([all]).
...
up_to_date

Prepare emodel

• Import Tensorflow Lite model

Use this python script to convert a TF Lite mode to an emodel. Before using this script, make sure you already have the tflite python library, which can be generated by the FlatBuffers compiler.

python import_tflite.py ./examples/tiny_conv_graph.tflite ./examples/tiny_conv_graph.emodel

Compile emodel:

kdl:compile("./examples/tiny_conv_graph.emodel", a_proj_dir)

Two C files will be generated under a_proj_dir, main_demo.c and model.c. The whole model is defined in model.c.

Build model

Modify main_demo.c to feed in input data, then build the model:

gcc main_demo.c model.c