ImageClassification

Image Classification Sample App

This sample app classifies images. The top 3 predicted Imagenet classes are displayed.

The app aims to showcase best practices for using TF Lite for model inference on Android devices.

Prerequisites

1. Clone this repository with Git-LFS enabled.
2. Download Android Studio. Version 2023.1.1 or newer is required.
3. Enable USB debugging on your Android device.

Build the APK

1. Download or export a compatible model from AI Hub Models.
2. Copy the .tflite file to src/main/assets/<your_model>.tflite
3. In ../gradle.properties, modify the value of classification_tfLiteModelAsset to the name of your model file (<your_model>.tflite)
4. Open the PARENT folder (android) (NOT THIS FOLDER) in Android Studio, run gradle sync, and build the ImageClassification target.

Supported Hardware (TF Lite Delegates)

By default, this app supports the following hardware:

• Qualcomm Hexagon NPU -- via QNN
• GPU -- via GPUv2
• CPU -- via XNNPack

Comments have been left in TFLiteHelpers.java and AIHubDefaults.java to guide you on how to add support for additional TF Lite delegates that could target other hardware.

AI Model Requirements

Model Runtime Formats

• TensorFlow Lite (.tflite)

I/O Specification

INPUT	Description	Shape	Data Type
Image	An RGB image	[1, Height, Width, 3]	float32 or uint8 (zero point of 0, scale of 1 / 255)

OUTPUT	Description	Shape	Data Type
Classes	Imagenet Classes	[ 1, 1000 ]	float32, uint8, or int8 (any quantization params)

Compatible AI Hub Models

The below is a non-exhaustive list of AI Hub Models that should be compatible.

Not every model has been individually tested with this app. Please file an issue or reach out on Slack if you find a model in this list with app compatibility issues.

• ConvNext-Tiny
• ConvNext-Tiny-w8a8-Quantized
• DenseNet-121
• EfficientNet-B0
• GoogLeNet
• GoogLeNetQuantized
• Inception-v3
• Inception-v3-Quantized
• MNASNet05
• MobileNet-v2
• MobileNet-v2-Quantized
• MobileNet-v3-Large
• MobileNet-v3-Large-Quantized
• MobileNet-v3-Small
• RegNet
• RegNetQuantized
• ResNet101
• ResNet101Quantized
• ResNet18
• ResNet18Quantized
• ResNet50
• ResNet50Quantized
• ResNeXt101
• ResNeXt101Quantized
• ResNeXt50
• ResNeXt50Quantized
• ShuffleNet-v2
• ShuffleNet-v2Quantized
• SqueezeNet-1_1
• SqueezeNet-1_1Quantized
• Swin-Base
• Swin-Small
• Swin-Tiny
• VIT
• WideResNet50
• WideResNet50-Quantized

Replicating an AI Hub Profile / Inference Job

Each AI Hub profile or inference job, once completed, will contain a Runtime Configuration section.

Modify TFLiteHelpers.java according to the runtime configuration applied to the job. Comment stubs are included to help guide you (search for TO REPLICATE AN AI HUB JOB...)

Note that if your job uses delegates other than QNN NPU, GPUv2, and TFLite, then you will also need to add support for those delegates to the app.

Technologies Used by this App

• Android SDK
• TensorFlow Lite
• TF Lite Support Library
• QNN SDK (TF Lite Delegate)
• GPUv2 Delegate
• XNNPack Delegate

License

This app is released under the BSD-3 License found at the root of this repository.

All models from AI Hub Models are released under separate license(s). Refer to the AI Hub Models repository for details on each model.

The QNN SDK dependency is also released under a separate license. Please refer to the LICENSE file downloaded with the SDK for details.