msf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML

For more information on this work, please read our preprint on alphaXiv or arXiv.

Reproduce analytical results

1. install python packages using requirement.txt
2. Run python analysis_optimization.py. Please check the following parameters:

usage: analysis_optimization.py [-h] [-m M] [-p P] [-c1 C1] [-c2 C2] [--log LOG]

options:
  -h, --help  show this help message and exit
  -m M        Model: MBV2-w0.35; MN2-vww5; MN2-320K
  -p P        Problem: [P1, P2, vanilla, heuristic]
  -c1 C1      specify the constraint set for the optimization problem P1: F_max. format: n1, n2...n_N
  -c2 C2      specify the constraint set for the optimization problem P2: P_max in kB. format: n1, n2...n_N
  --log LOG   Log Level

Reproduce Experimental Results on MCU

The experiments were conducted via RIOT-ML. To reproduce the results it is required to get RIOT-ML ready.

0. Go get a IoT board e.g. STM32 nucleo-f767zi. Connect the IoT board to your PC.
1. Clone the RIOT-ML from https://github.com/TinyPART/RIOT-ML/.
2. Follow the Prequisites section, install all necessary packages and toolchains.
3. Copy msf_cnn_eval_hardware.py, main.c and msf_CNN_eval_models to the RIOT-ML directory.
4. Run python msf_cnn_eval_hardware.py under the RIOT-ML directory.
5. Grab some coffee and wait for the results written in msf_cnn_HIL_eval_result_{board}.json :).