TF deploy / Rust

A tiny TensorFlow inference-only executor.

Why ?

TensorFlow is a big beast. It is designed for being as efficient as possible when training NN models on big platforms, with as much support for custom hardware as practical.

Performing inference, aka running trained models sometimes needs to happen on small-ish devices, mobiles phones and stuff. Cross-compiling TensorFlow for these platforms can be a daunting task, and produces huge libraries.

This project started as a very pragmatic answer to a critical problem we encountered at Snips recently: we needed to run (tiny) model as part of a library that we were porting to Android. The inference-only C interface that we were relying on other platforms (libtensorflow) was not provided nor buildable for Android. We wasted so much time trying that we decided we needed another option and we started this project on the side as a plan B.

It turns out we finally managed to build libtensorflow in time. As a matter of fact, TensorFlow team released their own Android build scripts just a few days after we managed to craft ours.

So this project is only a hobby of mine right now.

Status

This is very far to support any arbitrary model. Right now, we have a skeleton interpreter, and only a handful of naive implementation for actual Ops. Just what we needed for Google's Inception v3 to run. Moreover, only the strictly necessary data types have been implemented (most operators right now only operate on f32, a handful on integers).

Adding an Op is relatively straightforward, adding a data type more complicated.

BLAS backends and performance evaluation

Two features are provided: accelerate and openblas. They will plug BLAS backends into ndarray. Execution will be faster, to the price of portability.

This is a highly unscientific bench, performed on one single datapoint. I timed Inception v3 running on Grace Hopper image (not that the actual data is supposed to make a difference). It was run on my laptop (a mid-2014 MacBook pro).

• tensorflow reference (brew version, generic x86_64): 408ms
• tensorflow reference (optimized for my laptop): 323ms
• tensorflow deploy, no BLAS: 890ms
• tensorflow deploy, OpenBlas: 567ms
• tensorflow deploy, Accelerate: 544ms

Roadmap

One important guiding cross-concern: I want this library to cross-compile as easily as practical to small-ish devices (think 30$ boards).

• cleanup and generalize (op-wise and type-wise) basic operators (arithmetic, shape). consider factorizing paramater reading code and datatype switching
• find and integrate other TF models to use as example, test and bench
• investigate alternative impls for Conv2D
• refactor interpreter: make it stack-based (because it's easy), stop cloning everything
• consider ops accepting borrowed matrixes when it makes sense to avoid more clones
• consider having a separate set of non-TF mimicking operators
• optimise some ops combination (mul followed by add -> GEMM for instance)

License

Note: files in the protos directory are copied from the TensorFlow project and are not covered by the following licence statement.

Apache 2.0/MIT

All original work licensed under either of * Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0) * MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT) at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.