How to train custom models?

Hi, new to tfjs, just tested out tfjs-node mobilenet with pre-trained model. I would like to learn how to train custom models using exiting video footage. I came across roboflow.com which looks like to give me what I want, however it only offer formats in TFRecords and Tensorflow Object Detection. Any suggestions what tools I should use?

2 Likes

What specific frameworks are you looking for? Object Detection API is a well-developed and appreciated framework among the vision community. It comes with performance benefits, a rich repository of SOTA model implementations many of which are TFLite-compatible.

But here’s a standalone tutorial from keras.io that shows you how to train a RetinaNet from scratch:

1 Like

Hello there! Thanks for posting your TensorFlow.js question and welcome to the community :slight_smile: Indeed you may be interested in this codelab that shows you how to make your custom image classifier via a form of transfer learning using mobilenet as the base TensorFlow.js Transfer Learning Image Classifier  |  Google Codelabs

This is how the popular Teachable Machine (Teachable Machine) works behind the scenes to take video and retrain a network on top of mobile net to then recognize something new and all the retraining is done in < 30 seconds in the browser live if you have say 50 images of the new class or so.

If you wish to use MP4 files then you can simply take the frames of these if they only contain the new object you want to recognize and then use those frames as the training data.

1 Like