Conditional BW GAN on Arduino Uno (16x16)

Tiny class-conditional GAN that generates grayscale digit images directly on Arduino Uno and streams them to PC as image blocks for saving to PNG.

Highlights

• Runs on Arduino Uno (ATmega328P, 2KB SRAM / 32KB Flash).
• Class-conditional generation: request a specific digit 0..9.
• Grayscale output (16 levels) instead of binary-only output.
• End-to-end pipeline:
  1. Train cGAN on MNIST (Python/PyTorch)
  2. Quantize generator to INT8 and export Arduino header
  3. Run inference on Uno
  4. Capture serial output and save PNG files on PC
• Near-max Uno optimization:
  • Flash usage ~94% for higher-capacity model
  • SRAM remains safe for runtime buffers

Architecture

Generator (exported to Uno):

• Input: latent noise + one-hot class (10)
• MLP: 26 -> 48 -> 64 -> 256
• Conditional class embeddings injected at each layer
• Output: 16x16 image (grayscale quantized to 4-bit levels for protocol)

Discriminator (training only on PC):

• Conditional MLP using image + class one-hot.

Files

• gan_bw_arduino/train_and_export_bw_gan.py
  Train conditional GAN on MNIST and export INT8 header.
• gan_bw_arduino/tiny_bw_gan.ino
  Arduino Uno inference code (conditional generation + serial protocol).
• gan_bw_arduino/tiny_bw_gan_model.h
  Auto-generated quantized model for Uno.
• gan_bw_arduino/capture_bw_gan_images.py
  Reads serial protocol and saves PNG images.
• gan_bw_arduino/README.md
  Quickstart commands.

Quickstart

1) Train + export model

python3 gan_bw_arduino/train_and_export_bw_gan.py --epochs 30

2. Build + flash Uno

mkdir -p gan_bw_arduino_sketch
cp gan_bw_arduino/tiny_bw_gan.ino gan_bw_arduino_sketch/gan_bw_arduino_sketch.ino
cp gan_bw_arduino/tiny_bw_gan_model.h gan_bw_arduino_sketch/
arduino-cli compile --fqbn arduino:avr:uno gan_bw_arduino_sketch
arduino-cli upload -p /dev/cu.usbserial-A50285BI --fqbn arduino:avr:uno gan_bw_arduino_sketch

3. Generate PNGs for a chosen digit

python3 gan_bw_arduino/capture_bw_gan_images.py \
  --port /dev/cu.usbserial-A50285BI \
  --digit 7 \
  --count 20 \
  --out-dir gan_bw_arduino/generated_images_cgan7 \
  --scale 16