This repository contains the source-code and APK developed for our paper: Continual Learning on the Edge with TensorFlow Lite
- Offline Experiments includes the source-code necessary to reproduce the results presented in section "Enhancing TensorFlow Lite Capabilities with Continual Learning".
- Android Demo App includes the source-code of our Continual Learning application built on top off TensorFlow Lite.
- cldemo.apk can be used to install and test our demo-app on an Android smartphone device.
In order to compare their performance in real-world scenarios, both Transfer Learning and Continual Learning models were deployed on a Samsung Galaxy S10, Android device using TensorFlow Lite.
- A video demonstrating the training and inference during the on-device experiments for all scenarios can be seen here.
- A video demonstrating the training and testing of the CL model under non-ideal conditions can be seen here.
You can download the Android demo-app APK and install it on your own device. Compatible with Android 8.0 and later versions.
- To reproduce the results presented in section "Enhancing TensorFlow Lite Capabilities with Continual Learning", it is first necessary to download the CORe50 dataset.
- You also need to change the root in experiments.py to the path where your dataset is located.
- All experiments bellow measure accuracy over time based on the CORe50 NICv2 - 391 benchmark
- Compare the Transfer Learning model with the Continual Learning model
python controller.py --exp_tl_vs_cl - Compare FIFO with Random Replacement of old samples in the replay buffer
python controller.py --exp_sample_replacement - Compare different replay buffer sizes
python controller.py --exp_buffer_size