I really like the new model upload feature, can we please allow the ability to upload a tensorflowJS model? I would be willing to help with the code needed to convert a tensoflowJS model to tensorflowLite.
It does bring up the question of if EdgeImpulse should support the Javascript Machine Learning Community? My biased opinion is that Javascript Machine learning can increase the number of EdgeImpulse Enterprise customers without massively increasing the number of non-paying customers, for the simple fact that: Javascript can do most of the basics of what the EdgeImpulse client can do. See my demo below.
I did have a chance to briefly chat with @Alessandro at the ICTP workshop in Italy on Widening Access to TinyML Network by Establishing Best Practices in Education | (smr 3851) where I am presenting a method to simplify the making of CSV files for microcontroller sensor data using webSerial on a single web page. My main live webpage demo is here webpage index here, github here the demo uses the Arduino Nano33BleSense (Rev1, code avalable for Rev2). My video tutorial about it is tinyMLjs-intro - YouTube
.
.
My TFConverter command line method of converting TensorflowJS models to Arduino is as follows at this repository here:
Convert TensorflowJS to Keras
tensorflowjs_converter --input_format=tfjs_layers_model --output_format=keras_saved_model ./model.json ./
Convert Keras to TensorflowLite
tflite_convert --keras_model_file ./ --output_file ./model.tflite
Convert TensorflowLite to C-header file
xxd -i model.tflite model.h
The problem with the above code that EdgeImpulse could massively help with, is that many Javascript users are not comfortable with loading a docker to work with the command line or installing Python to do the conversion. Why Google has not made a pure Javascript model converter is beyond me.
Thank you for reading this feature request.
.
.
P.S. I have an old feature request to allow the export of TensorflowJS models from EdgeImpulse here Feature Request: export TensorflowJS
