1 # TensorFlow Lite 2019 Roadmap 2 3 **Updated: March 6th, 2019** 4 5 The following represents a high level overview of our 2019 plan. You should be 6 conscious that this roadmap may change at anytime relative to a range of factors 7 and the order below does not reflect any type of priority. As a matter of 8 principle, we typically prioritize issues that the majority of our users are 9 asking for and so this list fundamentally reflects that. 10 11 We break our roadmap into four key segments: usability, performance, 12 optimization and portability. We strongly encourage you to comment on our 13 roadmap and provide us feedback in the TF Lite discussion groups and forums. 14 15 ## Usability 16 17 * **More ops coverage** 18 * Prioritize many more ops based on user feedback 19 * **Op versioning & signatures** 20 * Op kernels will get version numbers 21 * Op kernels will be identifiable by signature 22 * **New Convertor** 23 * Implementing a new TensorFlow Lite convertor that will better handle 24 graph conversion (i.e. control flow, conditionals etc) and replace TOCO 25 * **Continue to improve TF Select Ops** 26 * Support more types of conversion utilizing TF Selects such as hash 27 tables, strings etc. 28 * Support smaller binary size when using select TF ops via op stripping 29 * **LSTM / RNN support** 30 * Add full support of conversion for LSTMs and RNNs 31 * **Graph Visualization Tooling** 32 * Provide enhanced graph visualization tooling 33 * **Pre-and-post processing support** 34 * Add more support for pre-and-post processing of inference 35 * **Control Flow & Training on-device** 36 * Add support for control flow related ops 37 * Add support for training on-device 38 * **New APIs** 39 * New C API as core for language bindings and most clients 40 * Objective-C API for iOS 41 * SWIFT API for iOS 42 * Updated Java API for Android 43 * C# Unity language bindings 44 * **Add more Models** 45 * Add more models to the support section of the site 46 47 ## Performance 48 49 * **More hardware delegates** 50 * Add support for more hardware delegates 51 * **Support NN API** 52 * Continually support and improve support for NN API 53 * **Framework Extensibility** 54 * Enable simplistic overwriting of CPU kernels with customized optimized 55 versions 56 * **GPU Delegate** 57 * Continue to extend the total support ops for OpenGL and Metal ops 58 * Open-source 59 * **Improve TFLite CPU performance** 60 * Optimizations for float and quantized models 61 62 ## Optimization 63 64 * **Model Optimization Toolkit** 65 * Post training quantization + hybrid kernels 66 * Post Training quantization + fixed-point kernels 67 * Training with quantization 68 * **More support for more techniques** 69 * RNN Support 70 * Sparsity/Pruning 71 * Lower bit-width support 72 73 ## Portability 74 75 * **Microcontroller Support** 76 * Add support for a range of 8-bit, 16-bit and 32-bit MCU architecture use 77 cases for Speech and Image Classification 78