TensorFlow Lite Android Benchmark

Set of tools to benchmark performance of TensorFlow image classification networks on Android devices.

Folder android/ contains Android app project. Folder scripts/ contains Python scripts for converting existing TensorFlow models to TensorFlow Lite models embeddable into the Android application.

Extends official TensorFlow Android example and adds new models and functionality to benchmark performance.

Generating example networks

Use script generate.py

Generates 6 networks pre-trained with ImageNet and saves them to h5 format

• NASNetMobile with input image size 224x224
• EfficientNetB0 with input image size 224x224
• MobileNetV2 with input image size 96x96
• MobileNetV2 with input image size 128x128
• MobileNetV2 with input image size 160x160
• MobileNetV2 with input image size 224x224

Converting TensorFlow model in h5 format to TensorFlow Lite

Convert to TFLite format

Use script convert.py

Converts TensorFlow model in h5 format to TensorFlow Lite format with optional dynamic range quantization

Usage: python convert.py -i <input file> [-o <output file>] [-q <quantization level>] [-s <samples count>]

Quantization levels:

• -q 0 - no quantization (default)
• -q 1 - dynamic range quantiation
• -q 2 - full integer quantization
  • this quantization level requires model to be calibrated using example input images, use -s parameter to control how many sample images are used for the calibration

(optional) Evaluate TF Lite model

Use script interpret.py

Evaluates TFLite image classification model Usage: python interpret.py -m <path to model> [-s <samples count>]

Insert metadata into model

Use script setMetadata.py

Inserts metadata into tflite model Usage: python setMetadata.py, then enter information after prompted by terminal

(optional) Verify model metadata

Use script getMetadata.py

Prints metadata from TFLite model file Usage: python getMetadata.py -i <input .tflite file>

Default naming conventions

NetworkName_InputSize_QuantizationLevel_HasMetadata

Example:

1. generator.py generates MobileNetV2_224.h5 - MobileNetV2 with input size of 224x224
2. convert.py quantizatizes and converts this model to TF Lite format and appends Q followed by quantization level to filename MobileNetV2_224_Q1.tflite - quantization level of 1 (see quantization levels above)
3. setMetadata.py adds metadata to the model and appends _M to filename MobileNetV2_224_Q1_M.tflite