Q: I'm unable to clone the ML embedded evaluation kit. When I run the command git clone "ssh://review.mlplatform.org:29418/ml/ethos-u/ml-embedded-evaluation-kit" and get a permission denied (publickey) error. What should I do to get the code base?
A: When cloning the repository, it's recommended to use https protocol command instead of ssh, use:
git clone "https://review.mlplatform.org/ml/ethos-u/ml-embedded-evaluation-kit"
A good starting point to explore the repository is the quick starting guide.
Q: I’m running through the quick-start guide and I’m running into an error with pip. When I run ./build_default.py, I get an error ImportError: No module named pip, but pip is installed on my machine.
A: Network or third party repository issues can cause the build_default script to fail and leave build environment in a broken inconsistent state. Removing the build and resources_downloaded folders and running the script again may help. If the problem persist contact your Arm representative or open a discussion on https://discuss.mlplatform.org/c/ml-embedded-evaluation-kit.
Q: When pointing to the TensorFlow Lite file explicitly in the cmake command, I get the following error message:
CMake Error at scripts/cmake/util_functions.cmake:73 (message): Invalid file path. Description: NN models file to be used in the evaluation application. Model files must be in tflite format.
A: This issue is usually caused by an incorrect path to the model file, pointed by the -D<use_case>_MODEL_TFLITE_PATH parameter. Check that the path is correct, clean the build folder and re-run the cmake command.
Q: How can we interpret the NPU and CPU cycles in terms of latency? Is the latency a summation of the total cycles (idle and active NPU, active CPU)?
A: For Fast Model simulations, active NPU cycles should be representative of a real system. However, when running code samples on Corstone-300 FVP, active CPU cycles should not be used for any performance analysis or interpretation. The Cortex-M part of the Fast Model is not cycle accurate or approximate, meanwhile NPU (Ethos-U) part is cycle approximate. If you need to interpret cycles for Cortex-M part, you need to use FPGA system (based on MPS3) or cycle accurate modelling environment.
Q: I changed the model for a use case but when running the application, tensor allocation fails with an error that an op for a builtin opcode could not be found and the application failed to get registration from that op code.
A: If you change the default model for a use case to a new model then you should update the corresponding <model_name>Model.cc source file to include any new operators that this model has.
For example, if you changed the img_class model you would need to modify the arm::app::MobileNetModel::EnlistOperations() function in source/use_case/img_class/src/MobileNetModel.cc to enlist any new operators in the model.
Don't forget to also increment the ms_maxOpCnt variable in the corresponding header file (e.g. source/use_case/img_class/include/MobileNetModel.hpp if we are changing img_class).