compilation

On-Device Training Under 256KB Memory

This section shows how we translate a pytorch models into intermediate representations (IR), and how to transform this forward IR to training version and perform preprocessing.

Pre-requests:

1. Setup forked tvm third_party/tvm-hack
   • Pull the remove folder via git submodule update --init --recursive
   • Install GCC/Clang, LLVM (<14.0) and Python (<3.10)
   • Compile following tvm compile-from-source (enable LLVM during compilation)
   • Export the compiled TVM in path

     export TVM_HOME=<DIR to third_party/tvm-hack>
     export PYTHONPATH=$TVM_HOME/python:${PYTHONPATH}
     

2. Translate pytorch models into training IR

   python mcu_ir_gen.py 
   

   and it will generate all required IRs and information under ir_zoos.
3. Convert the IR into a json format to enable MCU integration

   python ir2json.py <target IR path>
   

   and the json files will be stored in current directory.