l05_tflite_delegate

TensorFlow Lite delegates

The aim of this laboratory is to enhance a simple TensorFlow Lite delegate.

Tasks

• Check the delegate-example subdirectory - go through the CMakeLists.txt and sources.
• Build the delegate example based on delegate's README.md.
• Check the delegate_experiment script.
• [1pt] Implement convert_onnx_to_tensorflow method - it should use onnx2tf.convert method. Yielded model can be saved using save method.
• [1pt] Implement convert_to_tflite method - it should load the model saved to file in convert_onnx_to_tensorflow. One method to load can be tf.keras.models.load_model, depending on used format.
• [1pt] Add loading delegate in the delegate_experiment - use --delegate-path argument.
• Run the test script using models/test-delegate-two-inputs.onnx model, i.e.:

python3 -m dl_in_iot_course.l05_tflite_delegate.delegate_experiment \
    --input-onnx-model-path models/test-delegate-two-inputs.onnx \
    --tensorflow-model-path build/test-delegate-model.pb \
    --compiled-model-path build/test-delegate.tflite \
    --delegate-path ./delegate-example/build/libeigen-delegate.so

• [1pt] Add Eigen-based implementation of addition in the delegate_experiment - use Eigen::Map and Eigen::VectorXf classes to implement it (remember to rebuild the C++ project after modifications).
• [1pt] Make sure the delegate_experiment script runs properly and doesn't fail using the delegate_experiment script (there should be no assert failure or any other error).

Additional factors:

• [1pt] Git history quality

Resources

• Implementing a Custom Delegate (TensorFlow Lite documentation)
• Eigen documentation