Setting up Windows 10

Setup for training

pip is a nightmare to use so instead use conda, it's fully automated and cpu inference is over 10x faster. https://towardsdatascience.com/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc

The summary:

• install miniconda https://repo.continuum.io/miniconda/Miniconda3-4.6.14-Windows-x86_64.exe or anaconda https://repo.anaconda.com/archive/Anaconda3-2019.03-Windows-x86_64.exe
• launch Anaconda Prompt
• conda create -y --name tf_gpu tensorflow-gpu & conda activate tf_gpu & conda install -y keras & conda install -y matplotlib & conda install -y pandas & conda install -y pillow
• install Visual Code https://code.visualstudio.com/download
• setup Visual Code
  • create a folder and a test.py file
  • launch Visual Code from a conda prompt conda run -n tf_gpu code
  • open file with code from the context menu
  • there will be a message on the left pane of code, accept to open the folder, this will become your Code project
  • code will also suggest to install the python extension, do it
  • when this is done, in the lower left corner there is the python version and Conda environment, select ('tf_gpu':conda)

IDE and traing are now working. After all that, you restart and from that point onwards you won't have to launch code from conda prompt.

Convert .h5 to .kmodel

• download and extract the latest windows build https://github.com/kendryte/nncase/releases/ in a directory ./ncc
• open a conda prompt and launch 'tflite_convert --keras_model_file .\avoidance.h5 --output_file .\avoidance.tflite' then .\ncc\ncc.exe -i tflite -o k210model --dataset .\training avoidance.tflite .\avoidance.kmodel