#triton #pre-compiled #forward #neural #neural-network #models #created

yanked esp-idf-triton

Embedded neural network forward feeding for precompiled models created by the triton_grow crate

7 releases

0.1.6 Jan 26, 2024
0.1.5 Jan 25, 2024

#8 in #triton

MIT license

45KB
952 lines

triton mascot icon

esp-idf-triton

Lightweight fork of the Triton crate meant for running precompiled models on embedded systems

crates.io Documentation

esp-idf-triton can be used to run precompiled ML models created by the main triton crate. Creating a model in a Triton program and using the network.serialize_triton_fmt("out.triton"); will create a .triton file that is compatible with being deserialized by esp-idf-triton.

File IO support is soon to come, but as a quick proof of concept esp-idf-triton can currently parse neural network strings copied from the .triton file and perform forwards predictions.

Example

let model_str = "D|3|2|10.845654 11.002682 -13.501029 -14.699452 -53.440483 -53.715294|
    -6.101849 49.06853 61.28852#D|1|3|30.350481 -78.40228 70.861206|-19.532055#D|1|1|15.161753|-3.7315714".to_string();

let mut xor_net = Network::deserialize_triton_fmt_string(model_str);

creates a simple Network based on a trained SIGMOID XoR predictor.

The .triton file format currently defaults to SIGMOID activations, but its a top priority to implement activation function support in this file format

If open source development is your thing, we at Triton would love additional work on anything that can be implemented, please contact eversonb@msoe.edu if you'd like to help out!

License

Licensed under the Apache License, Version 2.0 http://www.apache.org/licenses/LICENSE-2.0 or the MIT license http://opensource.org/licenses/MIT, at your option. This file may not be copied, modified, or distributed except according to those terms.

No runtime deps