Working example of incorporating AOT with Pb graph?

[kwotsin]

Is there a working example showing how AOT could compile a PB graph for use in android devices? I came to your repository following this slideshare: https://www.slideshare.net/Hadoop_Summit/optimizing-profiling-and-deploying-high-performance-spark-ml-and-tensorflow-ai-models-in-production-with-gpus

I find the possibility of compiling an optimized graph with AOT rather exciting - is there a way to do this?

[cfregly]

@kwotsin we haven't yet exposed a public version of this AOT functionality. I'll try to prioritize this for a near-term release as we have working code, but it's specific to one of our customers.

thanks for the issue. keeping this open for tracking purposes.

[kwotsin]

@cfregly Thanks for mentioning this! Great to hear that you have something working with AOT - how did you find the performance for AOT so far in terms of speed/accuracy/memory consumption (if you could share)?

[cfregly]

Relevant links:

• https://stackoverflow.com/questions/46556553/cannot-create-the-final-binary-using-tensorflow-aot-compilation-for-cpu-backend
• Error in creating the final binary using AOT compilation for CPU backend tensorflow/tensorflow#13482
• Tensorflow binary seems compiled to use SIMD instructions like AVX2 and FMA, but actually not? tensorflow/tensorflow#13500

[cfregly]

related to TensorFlow Lite: https://www.tensorflow.org/mobile/tflite/