Please star the repo if you find it useful…
TensorFlow Lite (TFLite) is a set of tools that help convert and optimize TensorFlow models to run on mobile and edge devices - currently running on more than 4 billion devices! With TensorFlow 2.x, you can train a model with tf.Keras, easily convert model to .tflite and deploy it; or you can download a pretrained TFLite model from the model zoo.
This is a curated list of TFLite models with sample apps, model zoo, helpful tools and learning resources. The purpose of this repo is to -
- showcase what the community has built with TFLite
- put all the samples side-by-side for easy references
- knowledge sharing and learning for everyone
Please submit a PR if you would like to contribute and follow the guidelines here.
Here are some new features recently announced at TensorFlow World:
- New MLIR-based TFLite converter - enables conversion of new classes of models such as Mask R-CNN and Mobile BERT etc, supports functional control flow and better error handling during conversion. It is now enabled by default in the nightly builds - see details in the updated & initial announcements.
- TFLite Android Support Library - documentation | Sample code (Android)
- Create your custom classification models easily with the TFLite Model Maker (
model customization API) - Colab tutorials for Image & Text - On-device training is finally here! Currently limited to transfer learning for image classification only but it's a great start - Blog | Sample code (Android). Here is an example from the community - on-device activity recognition for next-generation privacy-preserving personal informatics apps - Blog | Android. Leverage transfer learning for efficiently training context sensing models directly on the Android device without the need for sending data to the server.
- Accelerating TensorFlow Lite on Qualcomm Hexagon DSPs - Blog | Documentation
Here are the TFLite models with app / device implementations, and references.
Note: pretrained TFLite models from MediaPipe are included, which you can implement with or without MediaPipe.
| Task | Model | App | Reference | Source |
|---|---|---|---|
| Classification | MobileNetV1 (download) | Android | iOS | Raspberry Pi | Overview | tensorflow.org |
| Classification | MobileNetV2 | Recognize Flowers with TFLite on Android Codelab | Android | TensorFlow team |
| Classification | MobileNetV2 | Skin Lesion Detection Android | Community |
| Classification | EfficientNet-Lite0 (download) | Icon classifier Colab & Android | tutorial 1 | tutorial 2 | Community |
| Object detection | Quantized COCO SSD MobileNet v1 (download) | Android | iOS | Overview | tensorflow.org |
| Object detection | YOLO | Flutter | Paper | Community |
| Object detection | MobileNetV2 SSD (download) | Reference | MediaPipe |
| License Plate detection | SSD MobileNet (download) | Flutter | Community |
| Face detection | BlazeFace (download) | Paper | Model card | MediaPipe |
| Hand detection & tracking | Download: Palm detection, 2D hand landmark, 3D hand landmark |
Blog post | Model card | MediaPipe |
| Pose estimation | Posenet (download) | Android | iOS | Overview | tensorflow.org |
| Segmentation | DeepLab V3 (download) | Flutter | Paper | Community |
| Segmentation (Flutter Realtime) | DeepLab V3 (download) | Flutter | Paper | Community |
| Segmentation | DeepLab V3 (download) | Android | iOS | Overview | tensorflow.org |
| Hair Segmentation | Download | Paper | Model card | MediaPipe |
| Style transfer | Download: Style prediction, Style transform |
Overview | Android | tensorflow.org |
| Task | Model | App | Reference | Source |
|---|---|---|---|
| Question & Answer | DistilBERT | Android | Hugging Face |
| Text Generation | GPT-2 / DistilGPT2 | Android | Hugging Face |
| Text Classification | Download | Android | tensorflow.org |
| Text Classification | Download | iOS | Community |
| Text Classification | Download | Flutter | Community |
| Task | Model | App | Reference | Source |
|---|---|---|---|
| Speech Recognition | DeepSpeech | Reference | Mozilla |
These are TFLite models that could be implemented in apps and things:
- MobileNet- pretrained MobileNet v2 and v3 models.
- TFLite models
- TensorFlow Lite models with Android and iOS examples
- TensorFlow Lite hosted models with quantized and floating point variants
- TFLite models from TensorFlow Hub
These are TensorFlow models that could be converted to TFLite and then implemented in apps and things:
- Official TensorFlow models
- Tensorflow detection model zoo - pre-trained on COCO, KITTI, AVA v2.1, iNaturalist Species datasets
A list of ideas and projects - you can help by creating a tflite model ready for implementation, add a mobile app idea that needs a tflite model created, or write an end-to-end tutorial with sample code. This is also where you can seek help from the community.
- U-GAT-IT (Selfie <-> Anime) - project repo.
- Deeplab V3 - image segmentation, which is supported by TFLite but there’s no tutorial on how to convert Deeplab v3 TF models to TFLite.
- Mask-RCNN object detection, which is one of the most popular on-device ML use cases.
- DeepSpeech - a very popular ASR framework.
- Segmentation + Style Transfer - project repo.
- YOLO - overview
- Classify pose - overview.
- Sound classification - overview.
- SPICE (Pitch Detection) - overview.
- Speech Command - overview.
ML Kit is a mobile SDK that brings Google's ML expertise to mobile devs.
- 10/1/2019 ML Kit Translate demo with material design - recognize, identify Language and translate text from live camera with ML Kit for Firebase - Codelab | Android (Kotlin).
- 3/13/2019 Computer Vision with ML Kit - Flutter In Focus - tutorial.
- 2/9/219 Flutter + MLKit: Business Card Mail Extractor - tutorial | Flutter.
- 2/8/2019 From TensorFlow to ML Kit: Power your Android application with machine learning - slides | Android (Kotlin).
- 8/7/2018 Building a Custom Machine Learning Model on Android with TensorFlow Lite - tutorial.
- 7/20/2018 - ML Kit and Face Detection in Flutter - tutorial.
- 7/27/2018 ML Kit on Android 4: Landmark Detection - tutorial.
- 7/28/2018 ML Kit on Android 3: Barcode Scanning - tutorial.
- 5/31/2018 ML Kit on Android 2: Face Detection - tutorial.
- 5/22/2018 ML Kit on Android 1: Intro - tutorial.
- Edge Impulse - helps you to train TFLite models for embedded devices in the cloud. (@EdgeImpulse)
- Fritz.ai - an ML platform that makes iOS and Android developers’ life easier: with pre-trained ML models and end-to-end platform for building and deploying custom trained models. (@fritzlabs)
- MediaPipe - a cross platform (mobile, desktop and Edge TPUs) AI pipeline by Google AI. (PM Ming Yong) | MediaPipe examples
- Coral Edge TPU - Google’s edge hardware. Coral Edge TPU examples
- TFLite Flutter Plugin - provides a dart API similar to the TFLite Java API for accessing TensorFlow Lite interpreter and performing inference in flutter apps. tflite_flutter on pub.dev
- Netron - for visualizing models
- AI benchmark - for benchmarking computer vision models on smartphones
- Performance benchmarks for Android and iOS
- How to design machine learning powered features - material design guidelines for ML | ML Kit Showcase App
- The People + AI Guide book - learn how to design human-centered AI products
Interested but not sure how to get started? Here are some learning resources that will help you whether you are a beginner or a practitioner in the field for a while.
- 4/20/2020 - What’s new in TensorFlow Lite from DevSummit 2020, Khanh LeViet. (link)
- 4/17/2020 - Optimizing style transfer to run on mobile with TFLite, Khanh LeViet and Luiz Gustavo Martins. (link)
- 4/14/2020 - How TensorFlow Lite helps you from prototype to product, Khanh LeViet. (link)
- 11/8/2019 - Getting Started with ML on MCUs with TensorFlow, BRANDON SATROM. (link)
- 8/5/2019 - TensorFlow Model Optimization Toolkit — float16 quantization halves model size, the TensorFlow team. (link)
- 7/13/2018 - Training and serving a realtime mobile object detector in 30 minutes with Cloud TPUs, Sara Robinson, Aakanksha Chowdhery, and Jonathan Huang. (link)
- 6/11/2018 - Why the Future of Machine Learning is Tiny, Pete Warden. (link)
- 3/30/2018 - Using TensorFlow Lite on Android, Laurence Moroney. (link)
- 03/2020 - Raspberry Pi for Computer Vision (Complete Bundle | TOC) by the PyImageSearch Team: Adrian Rosebrock (@PyImageSearch), David Hoffman, Asbhishek Thanki, Sayak Paul (@RisingSayak), and David Mcduffee.
- 12/2019 - TinyML by Pete Warden (@petewarden) and Daniel Situnayake (@dansitu).
- 10/2019 - Practical Deep Learning for Cloud, Mobile, and Edge by Anirudh Koul (@AnirudhKoul), Siddha Ganju (@SiddhaGanju), and Meher Kasam (@MeherKasam).
- 4/1/2020 - Easy on-device ML from prototype to production (TF Dev Summit '20)
- 3/11/2020 - TensorFlow Lite: ML for mobile and IoT devices (TF Dev Summit '20)
- 10/31/2019 - Keynote - TensorFlow Lite: ML for mobile and IoT devices
- 10/31/2019 - TensorFlow Lite: Solution for running ML on-device
- 10/31/2019 - TensorFlow model optimization: Quantization and pruning
- 10/29/2019 - Inside TensorFlow: TensorFlow Lite
- 4/18/2018 - TensorFlow Lite for Android (Coding TensorFlow)
- Udacity Introduction to TensorFlow Lite - by Daniel Situnayake (@dansitu), Paige Bailey (@DynamicWebPaige), and Juan Delgado
- Coursera Device-based Models with TensorFlow Lite - by Laurence Moroney (@lmoroney)