MA_ZITI

Structure of repository

The repository contains scripts for running the algorithms presented in the master thesis of Hendrik Borras. The thesis itself can be found here: https://raw.githubusercontent.com/heborras/MA_ZITI/main/thesis/MA_HB_2021.pdf

Please be advised that opening the PDF through github's file browser may cash the JavaScript-PDF renderer of your browser. This will not happen, when opening the link above, which directly displays the PDF.

Automatic tuning of performance parameters, Chapter 4

The script implementing the algorithms of Chapter \ref{chap:tune} also interacts with parts of the Chapter on Pruning. Thus the script is designed to be run within the Docker container created by FINN. The script itself can be found here: https://github.com/heborras/MA_ZITI/tree/main/simd-pe-tuning

Pruning in FINN, Chapter 5

The script developed for the training of pruned networks with the iterative L^1-norm pruning method can be found here: https://github.com/heborras/MA_ZITI/tree/main/training

Setup and runnig scripts

Automatic tuning of performance parameters, Chapter 4

The jupyter notebook (https://github.com/heborras/MA_ZITI/blob/main/simd-pe-tuning/cnv_varied_bit_and_pruning_parallel_testing-0.4b_dev.ipynb) should be run from within the FINN docker container.

Setting up the required jupyter notebook server with FINN:

• First setup FINN using the steps described here: https://finn.readthedocs.io/en/latest/getting_started.html
• Then clone Hendrik Borras's fork of the FINN repository: https://github.com/heborras/finn
• Switch to the branch "feature/0.4_cutting_pruning".
• Copy the "cnv_varied_bit_and_pruning_parallel_testing-0.4b_dev.ipynb" notebook of this repository into the "notebooks" folder of the cloned FINN repository
• Run the jupyter notebook server in the feature branch from the command line: bash ./run-docker.sh notebook

Running the script:

• Open the script from jupyter notebook
• Modify settings in the script as required. Please try to understand these settings before running the script, independent of wether you changed something or not.
• Make sure to set the build_dir variable to a sensible path. This is where temporary models and folders will be created.
• Some settings are set in two external files. These influence how parallelization is realized during runtime and can even be modified during runtime to fine-tune performance.
  • build_dir + '/cpu_percentage_max.txt' Should contain one float between 1. and 99.. While the current CPU load is below this setting new optimization runs will be started. This allows to run multiple optimizations in parallel.
  • build_dir + '/num_workers.txt' Should contain one integer above 0. Here the number of workers per running optimization can be specified. This enables a second degree of parallelization.
  • Be careful what is set here. Large settings can lead to heavy CPU load and memory usage!
• Run the parallel implementation to start the optimization.

Results from the script:

• Running the script produces ziped JSON files, which contain information about the optimization and results from running the given networks on an Ultra96V2 FPGA.
• These ziped JSON files will be saved twice, once in the location specified by the build_dir variable and a second time in the root folder of the repository running the FINN docker container.

Pruning in FINN, Chapter 5

Setting up Brevitas and PyTorch:

• Install Anaconda: https://www.anaconda.com/
• Create a conda environment using the brevitas_torch-1-4.yml contained in this repository. The environment contains the Brevitas and PyTorch version used in the thesis work. The command could look something like this: conda env create -f brevitas_torch-1-4.yml
• Create some result json files using the script for automatically tuning the performance parameters and store them on the computer for training.
• Create a folder structure relative to the location of the script, where the resultin json files will be stored. Like this: mkdir -p finn_result_jsons/after_training/

Running the script:

• Activate the conda environment: conda activate brevitas_torch-1-4
• Run the script and give it the path to the ziped json file from script from chapter 4. As example like this: python Brevitas_train_pruning_from_FINN_json.py --finn_json finn_result.json.gz

Results from the script:

• During training the script will print out log data and store the log in a sub-folder, which is created when the script starts. Additionally, when the script ends the final model will be saved to this folder.
• At the end of the training the script will save the loss, accuracy and model information to the path: finn_result_jsons/after_training/ relative to its location.