Please consult here on how to install the tools.
In order to compile and run this example follow these steps:
xcore-opt --xcore-weights-file=model.params vww_quant.tflite -o model.tflite mv model.tflite.cpp model.tflite.h src xmake python -c 'from xmos_ai_tools import xformer as xf; xf.generate_flash( output_file="xcore_flash_binary.out", model_files=["model.tflite"], param_files=["model.params"] )' xflash --target XCORE-AI-EXPLORER --data xcore_flash_binary.out xrun --xscope bin/app_flash_single_model.xe
This should print:
No human (9%) Human (98%)
The difference with the version in ../app_no_flash
is that we have sent
the learned parameters into flash memory; this has significantly reduced
the size of the model. We can see this by looking at the size of the files:
% ls -l model.* -rw-r--r-- 1 henk staff 224576 18 Jul 11:07 model.params -rw-r--r-- 1 henk staff 20032 18 Jul 11:07 model.tflite
The model.params file needs to be made into a flash image, which is what
the python command does. Finally, before we execute it, we must program the
flash with the learned parameters, which is what xflash
is for.