@Vengineerの戯言 : Twitter
SystemVerilogの世界へようこそ、すべては、SystemC v0.9公開から始まった
TensorFlow Liteの新しいビデオ
さくっと、まとめてみました。
ruy/gemmlowp
ruyは、これ ですね。
ruy is a matrix multiplication library. Its focus is to cover the matrix multiplication needs of TensorFlow Lite.
ruy supports both floating-point (like Eigen) and quantized (like gemmlowp).
Acceleration (Delegates)
- Metal
TensorFlow Select
Select TensorFlow operators to use in TensorFlow Lite
Cross-platform bindings
Experimental APIs : C, C# (Unity), Objective C, Swift
Performance
Benchmarking : Model Benchmark tool, (Evaluation tool (+app), MLPerf)
Acceleration : Edge TPU, GPU, NN API, (CPU optimizations (arm => ruy & x86))
Optimization
Quantization : Legacy quantized training (CPU/NPU), Post-training quantization (CPU/NPU), (Keras-based quantized training (CPU/NPU))
”Hybrid” quantization (Post-training quantization with float & fixed point)
Integer quantization (Post-training quantization with fixed point math only)
Other optimization : Model optimization toolkit, Keras-based connection pruning, (Runtime sparsity support)
Documentation
Docs : New tensorflow.org/lite, (More tutorials)
Model repository : 5 models ready to use, (Expand repository to many more models)
Training
TF Select + Control Flow + Subgraphs + Variables
TensorFlow Lite Roadmap
New Converter :
Control flow (sooner)
Training (later)
Improved runtime
Idiomatic APIs for pre/post-processing
Stable bindings (C/Obj-C/Swift)
もう一つが、TensorFlow World 2019 Conference でのビデオ
上のビデオでは語られていないものだけ。
TensorFlow Lite Support Library
APIs for simplifying pre- and post-processing (launced)
Autogenerates pre- and post-processing (in progress)
Model Metadat support
.json Example
Language Bindings
Swift, Obj-C, C#, C for iOS, Android and Unity
Community language bindinsg : Rust, Go, Flutter/Dart
Running TensorFlow Lite on Microcontrollers
Our Future
More ops and supported models
Performance improvements (CPU, GPU, TPU, etc.)
More models, TF Hub integration & better documentation
On-device training and personalization
Automated performance discovery tools
