Workflow to efficiently compress and deploy DNN classifiers for SoC/FPGA

Efficient end-to-end workflow for deploying DNNs on an SoC/FPGA by integrating hyperparameter tuning through Bayesian optimization with an ensemble of compression techniques (quantization, pruning, and knowledge distillation).

The proposed workflow is based on three stages: DNN training and compression, integration with a hardware synthesis tool for ML, and hardware assessment.

DNN training and compression

Hardware assessment framework

Required libraries and tools

Libraries

Check the file requirements.txt inside the environment folder.

SoC/FPGA tools

• Vivado Design Suite - HLx Editions 2019.1, 2019.2, 2022.2

What's in this repository?

Repository tree:

• The workflow comprises the following folders:
  • 00-environment
  • 01-compressionAndTraining
  • 02-hls4mlIntegration
  • 03-assessmentFramework
  • 04-integrationPYNQ*

• *The folder integrationPYNQ is for those willing to integrate the ML IP core into the PYNQ framework.

Custom datasets

• Available under reasonable request.

Current Branches

Branch	Purpose
main	Clean refactor of original code
backup_original	Legacy structure linked to initial publication
full_compression	Combined pipeline for QAT + KD + pruning in a single loop
coming soon...	Separate branches for isolated quantization, pruning, and KD

---

Citation

If this codebase is linked to a publication, cite it as:

@ARTICLE{10360204,
  author={Molina, Romina Soledad and Morales, Iván René and Crespo, Maria Liz and Costa, Veronica Gil and Carrato, Sergio and Ramponi, Giovanni},
  journal={IEEE Embedded Systems Letters}, 
  title={An End-to-End Workflow to Efficiently Compress and Deploy DNN Classifiers on SoC/FPGA}, 
  year={2024},
  volume={16},
  number={3},
  pages={255-258},
  doi={10.1109/LES.2023.3343030}}

---

Acknowledgements

Built on top of:

• TensorFlow Model Optimization Toolkit
• QKeras
• Keras Tuner

---

Final remarks

Have fun!! And remember, this is a methodology to facilitate the training and compression process when targetting resource-constrained devices, it is not (yet ;) ) an automatic process.