12. Frequently Asked Questions (FAQs)ΒΆ

  1. Q: How to resolve the following error message when compiling an ONNX model exported from PyTorch?

    ERROR:TVMC: dynamic shape not supported.
    

    A: This error could be caused by exporting a PyTorch model with dynamic input and output shapes. For example, if your torch.onnx.export() call contains a dynamic_axes argument, remove it.

  2. Q: How to compile models in Keras(.h5) format via TFLite conversion for F29x devices?

    A: We recommend using TFLiteConverter utility in tensorflow python package to convert the model into TFLite format first, and then compiling the TFLite model.

    We do not recommend this path for non-F29x devices, for example, C28x/TI-NPU based devices, as we do not support TI-NPU QAT for Keras.

    # pip install tensorflow
    from tensorflow import lite
    from tensorflow import keras
    
    h5_model = keras.models.load_model("./model.h5")
    converter = lite.TFLiteConverter.from_keras_model(h5_model)
    tflite_model = converter.convert()
    with open("model.tflite", "wb") as f:
      f.write(tflite_model)