Hello everyone,
I’m a GSoC 2025 participant working this summer on the “Property Editor for Creating Transfer Learning ML Models” project.
My initial proposal was for the Color Choice Property Editor, but since I completed that early, my project was updated to property editor for creating transfer-learning ML models.
Abstract
The current Personal Image Classifier (PIC) and Personal Audio Classifier (PAC) components in MIT App Inventor depend on external websites for building and training models. This project aims to bring that functionality directly into the App Inventor interface via an embedded property editor.
Design Preview
Property Editor
This is the property editor for the Personal Image Classifier component, which opens a dialog for building your ML models.
Training Dialog
Users can define classes and add images using their webcam. They can also upload existing models and training data.
Training Tab
This tab allows users to set training parameters.
Current idea: Upon clicking "Start Training", the model will be trained using the uploaded data via TensorFlow.js.
Testing Tab
Users can test their trained model here. This also allows exporting the model and training data in JSON format.
Current idea: Once the model is created, it should automatically be uploaded to the project assets (replacing any older model if present).
Full Proposal
This is my current implementation plan and design flow. I kindly request your valuable feedback on the project. I’m particularly eager to hear your thoughts on the tab flow (Classifications → Train → Testing).
If I’ve missed anything or if you have suggestions for improvements, please share — this is the best time to make meaningful changes to enhance the project.
Best regards,
Sumit Kumar
