Skip to main content

Model Conversion

How to accelerate your model on ARTPEC-7​

  1. Before you begin, read the Optimization Tips for information that is useful for the following steps.

  2. Train your model using the framework of your choice. Common options are PyTorch or TensorFlow.

  3. Quantize (int8/uint8) and export your trained model to the .tflite format. How you do this depends on the framework you use. We provide guides for PyTorch and TensorFlow:
    PyTorch-to-tflite
    TensorFlow-to-tflite

  4. Use the Edge-TPU compiler to compile the .tflite model to an edge-tpu .tflite model. Refer to the documentation for how to do this.

  5. Run the edge-tpu .tflite model on your device. Example test run:

    larod-client -g <my-model>.tflite -c google-edge-tpu-tflite -i ""

    More info here

info

Note that not all ARTPEC-7 devices has an Edge-TPU. For devices without an Edge-TPU, you could possibly run the regular .tflite model (skip step 3), on the GPU or the CPU. The model will however run slower than it would on an Edge-TPU.

larod-client -g <my-model>.tflite -c axis-a7-gpu-tflite -i ""
larod-client -g <my-model>.tflite -c cpu-tflite -i ""

Flowchart​

DLPU Conversion Schematic