Tiny CNN - Developing Efficient CNN Architectures for Image Classification on Microcontroller Units

This repository contains the notebooks for the CNN-Analyzer - a dashboard based tool for determining optimal CNN architecture parameters

No.	File Name	Description
01	01 Model Factory.ipynb	tf.model.summary tf.keras.utils.plot_model Visualization of model graph Untrained TensorFlow model
02	02 TensorFlow Layer Analyzer.ipynb	tf.model.layers layerwise model info
03	03 Model Analyzer.ipynb	Parsing .txt MACs FLOPs Total parameters Trainable parameters Non-trainable parameters Create dataframe from _layers.csv Untrained TensorFlow model conversion to TFLite TFLite model Untrained TensorFlow model conversion to TFLite with int8-quantization Int8 quantized TFLite model TFLite model size in KB TFLite int8 model size in KB tflite_tools.py (E. Liberis) layerwise RAM in bytes Max value from layerwise RAM in bytes General model architecture parameters Model name Model variation Alpha factor Image resolution Number of classes Number of input channels (RGB or grayscale) MLTK Profiler (hardware simulation) layerwise metrics (ops, MACs, CPU cycles, energy, input shape and output shape) Model size in bytes RAM in bytes FLOPs MACs Number of unsupported layers Int8 energy
04	04 Deploy to MCU.ipynb	Convert TFLite model to C code
05	05 LINUX - Benchmarking TFLite Models.ipynb	TensorFlow native benchmarking tool Layerwise benchmarking report of int8 model running on desktop
06	06 LINUX - Parsing CPU Benchmark Data.ipynb	Parsing _benchmark.txt
07	07 Model Training.ipynb	Trained TensorFlow model TFLite model Int8 quantized TFLite model
08	08 Update Model DB.ipynb	Write test accuracy from model training project to the model database
09	09 Benchmarking on STM32-Cube.AI.ipynb	Requires profiling int8 quantized with STM32-Cube.AI on external MCU Parsing <model_name>_stm32_benchmark/network_validate_report_INT8.txt CPU cycles per MAC Inference time in milliseconds Inference duration and layer type per layer
10	10 Model Comparison & Visualizations.ipynb	Create visualizations

Install the Conda Environment

conda env create -f conda-environment.yhttps://github.com/subrockmann/tiny_cnn/blob/master/08%20Update%20Model%20DB.ipynb

Activate the Environmnet

conda activate tiny_cnn

Update the Environment

conda activate tiny_cnn
conda env update --file environment.yml # --prune

Uninstalling the Environment

If there are too many error about inconsistencies, it might be easier to uninstall the environment and re-install it from scratch.

conda deactivate
conda env remove --name tiny_cnn

If Nothing Else Helps

conda clean --all

You Might Want to Run:

conda update --all

Libcupti Issue:

conda install -c nvidia cuda-cupti

Setting up the Arduino IDE

Unfortunately the Arduino_TensorFlowLite library is not include in the Arduino Library Manager. Instead the library has to be installed from a ZIP File https://www.ardu-badge.com/Arduino_TensorFlowLite/zip