Running MoViNet-Stream Models on Raspberry Pi with TFLite

1. Download Code Example

git clone https://github.com/berlin0308/Raspberrypi-MoViNet-TFLite.git --depth 1

2. Install Python Packages

sh setup.sh

• For model inference on Raspberry Pi, install tflite-runtime only instead of the whole tensorflow package

3. Run Inference

python3 classify.py --model models/a1_v2_stream_b16_lr0.0001_g1.5_d0.5_918/a1_float16.tflite

• --model : MoViNet model path
• --label : Path to the label map txt
• --numThreads : CPU threads to run the model

Model Details

Model ID	Quantization	Model Path (* Recommended)	Latency	F1-score
A0	Int8	models/a0_v5_stream_b16_lr0.0004_g2.0_d0.1_sd0.0_92/a0_int8.tflite	35 ms	0.312
A0	Float16	models/a0_v5_stream_b16_lr0.0004_g2.0_d0.1_sd0.0_92/a0_float16.tflite	47 ms	0.916
A1	Int8	models/a1_stream_b16_lr0.0001_g1.5_d0.5_918/a1_int8.tflite	49 ms	0.579
A1	Float16	models/a1_stream_b16_lr0.0001_g1.5_d0.5_918/a1_float16.tflite	98 ms	0.918
A2	Int8	models/a2_stream_b16_lr0.0004_g1.2_d0.5_sd0.3_929/a2_int8.tflite	78 ms	0.764
A2	Float16	models/a2_stream_b16_lr0.0004_g1.2_d0.5_sd0.3_929/a2_float16.tflite	154 ms	0.929

• Raspberry Pi 4 Model B
• Raspberry Pi OS / 64-bit / 4 Threads