Some question about EON Compiler

baozhu1 · January 21, 2022, 6:05am

Hi Edge Impulse team,
I checked this code. and compare with tflite_micro_compiler generated C code. I found they are mostly the same but some debug info.

github.com

edgeimpulse/firmware-himax-we-i-plus/blob/master/tflite-model/trained_model_compiled.cpp

/* Generated by Edge Impulse
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/

This file has been truncated. show original

Anyone can talk more about EON Compiler, If these tools are based on Open Source code. can we send PR to the project and make it better?

janjongboom · January 21, 2022, 8:38am

Hi @baozhu1 EON Compiler is indeed based on that project, but has a bunch of extra things, like support for extra memory allocation methods, support for custom ops, support for CMSIS-NN arena size calculation, and actually supporting latest TensorFlow Lite models / kernels (upstream project is stuck on TF2.3).

It’s currently not open source though, but would be happy to hear any suggestions.

Have cleared up the edgeimpulse/tflite_micro_compiler repository, that is no longer relevant.

baozhu1 · January 21, 2022, 11:14am

@janjongboom The tools are written via C++, It’s better rewritten using Python, and makes it live longer.

janjongboom · January 21, 2022, 11:31am

At least if we write it in C++ I know it will still compile in 10 years, unlike the Python dependency hell we encounter on a daily basis

On a more serious note, C++ is probably the right choice here because it’s easy to get things like the required memory allocation from tensors. You can just instantiate them, and then see how much they’ve allocated. That part will always be native, so more straight forward to write everything in C++.

VikramDattu · March 9, 2022, 11:39am

Hi @janjongboom Vikram here from Espressif Systems. Would like to add some changes to the tflite_compiler to make it more specific to ESP32 and ESP32-S3 etc to sqeeze more performance.

As you mentioned, the github tflite_compiler is stuck to older version. It has r2.4 branch as well which is bit later than tensorflow 2.3 but still much older. Is it possible to upstream changes from EON to this repo, so that it is easy to maintain and contribute by broader audience?
If that’s not in the plans, is there a way to collaborate on this front? This will help to avoid rework from my side to rework from my side to update the repo.

Espressif’s repo for your reference: https://github.com/espressif/tflite-micro-esp-examples

jefffhaynes · April 10, 2022, 2:59pm

FWIW your deployment page seems to imply that the EON compiler is open source. Might be good to clarify.