- tnesorflow的瘦身版本 - 主要用於推論
- 基本上會在tensorflow.lite模組中，主要會使用到的有Interpreter以及TFLiteConverter
- 官方說明 - ML for Mobile and Edge Devices - TensorFlow Lite
- 可以看到介紹，stackoverflow，以及usecase
- 官方文件 Get started with TensorFlow Lite
- TensorFlow Lite converter - 例子很少，加減看
- TensorFlow Lite and TensorFlow operator compatibility
- code-base basically C++
- TensorFlow Lite概述：转换器、解释器、XLA和2019年路线图 贊同80+, 2019, July
-
可以將
tflite模型compile成edgetpu support的形式，那麼就可以使用edgetpu做推論，需要使用edgetpu compiler，posenet中的BasicEngine有python接口 -
can edge TPU run two model at the same time?
- [Background] yes, but it might be slow, because edgetpu cache the model parameters in edgetpu memory, which enabling fast inference speed. when running a second model if we want, requires swapping the model parameter data in RAM，
- check official doc Run multiple models with multiple Edge TPUs
- [Possible solution] co-compiling your models, which allows the Edge TPU to store the parameter data for multuple models in RAM together, which means it typically works well onlu for small models
- [Background] yes, but it might be slow, because edgetpu cache the model parameters in edgetpu memory, which enabling fast inference speed. when running a second model if we want, requires swapping the model parameter data in RAM，
-
若API flow的推論速度或推論精確度不符合預期，那麼就需要考慮自己做model puring，distilling，參考 README中的 Model Compression章節
-
keras / tensorflow(pb file) ->tflite(.tflite) ->edgetpu_compiler(.tflite)- example posenet_mobilenet_v1_edgetpu.tflite 1.3M