BitNetMCU: High Accuracy Low-Bit Quantized Neural Networks on a low-end Microcontroller

BitNetMCU is a project focused on the training and inference of low-bit quantized neural networks, specifically designed to run efficiently on low-end microcontrollers like the CH32V003. Quantization aware training (QAT) and fine-tuning of model structure and inference code allowed surpassing 99% Test accuracy on a 16x16 MNIST dataset without using multiplication instructions and in only 2kb of RAM and 16kb of Flash.

Update Introducing a new model architecture based on deep separable convolutions allowed to push the accuracy even further to 99.55% accuracy, meeting state-of-the-art MNIST accuracy for CNNs while still fitting into the same memory constraints. This model requires a hardware multiplier, which is available in many low-end RISC-V and ARM Cortex-M0 microcontrollers.

The training pipeline is based on PyTorch and should run anywhere. The inference engine is implemented in Ansi-C and can be easily ported to any Microcontroller.

You can my buildlogs on the project in the docs/ directory here and CNN details here. Latest update on optimize fc here. Also see BitNetMCU blog articles

Project Structure

BitNetMCU/
│
├── docs/                      # Report
├── mcu/                       # MCU specific code for CH32V003
├── modeldata/                 # Pre-trained models
│
├── BitNetMCU.py               # Pytorch model and QAT classes
├── BitNetMCU_inference.c      # C code for inference 
├── BitNetMCU_inference.h      # Header file for C inference code
├── BitNetMCU_MNIST_test.c     # Test script for MNIST dataset
├── BitNetMCU_MNIST_test_data.h# MNIST test data in header format (generated)
├── BitNetMCU_model.h          # Model data in C header format (generated)
├── exportquant.py             # Script to convert trained model to quantized format
├── test_inference.py          # Script to test C implementation of inference
├── training.py                # Training script for the neural network
└── trainingparameters.yaml    # Configuration file for training parameters

Training Pipeline

The data pipeline is split into several Python scripts for flexibility:

1. Configuration: Modify trainingparameters.yaml to set all hyperparameters for training the model.

2. Training the Model: The training.py script is used to train the model and store the weights as a .pth file in the modeldata/ folder. The model weights are still in float format at this stage, as they are quantized on-the-fly during training.

3. Exporting the Quantized Model: The exportquant.py script is used to convert the model into a quantized format. The quantized model weights are exported to the C header file BitNetMCU_model.h.

4. Optional: Testing the C-Model: Compile and execute BitNetMCU_MNIST_test.c to test inference of ten digits. The model data is included from BitNetMCU_MNIST_test_data.h, and the test data is included from the BitNetMCU_MNIST_test_data.h file.

5. Optional: Verification C vs Python Model on full dataset: The inference code, along with the model data, is compiled into a DLL. The test-inference.py script calls the DLL and compares the results with the original Python model. This allows for an accurate comparison to the entire MNIST test data set of 10,000 images.

6. Optional: Testing inference on the MCU: follow the instructions in mcu/readme.md. Porting to architectures other than CH32V003 is straightforward and the files in the mcu directory can serve as a reference.

Updates

• 24th April 2024 - First release with Binary, Ternary, 2 bit, 4 bit and 8 bit quantization.
• 2nd May 2024 - tagged version 0.1a
• 8th May 2024 - Added FP1.3.0 Quantization to allow fully multiplication-free inference with 98.9% accuracy.
• 11th May 2024 - Fixes for Linux. Thanks to @donn
• 19th May 2024 - Add support for non-symmetric 4bit quantization scheme that allows for easier inference on MCUs with multiplier. The inference code will now use code optimized for multiplierless MCUs only on RV32 architectures without multiplier.
• 20th May 2024 - Added quantscale as a hyperparameter to influence weight scaling. Updated documentation on new quantization schemes.
• 26th May 2024 - tagged version 0.2a
• 19th July 2024 - Added octav algorithm to calculate optimal clipping and quantization parameters.
• 26th July 2024 - Added support for NormalFloat4 (NF4) Quantization. Updated documentation
• 7th September 2025 - New CNN architecture based on sequential depthwise separable convolutions allows to reach 99.55% accuracy while still fitting into 16kb Flash and 4kb RAM. See documentation for details.
• 2nd January 2026 - Add export and inference code for ternary quantization. See documentation for details.
• 6th September 2026 - Optimization of fc models, achieving 99.37% with Muon, NF4A4 / S4A8, activiation quantization and patch-wise processing. See documentation.