TinyDigitNet

A 14,410-parameter CNN for 28x28 grayscale digits, sized to run under ExecuTorch on Zephyr RTOS boards with no delegate: it has been run on an RP2040 (Cortex-M0+, Arducam Pico4ML-BLE, reading digits from the on-board camera), an ESP32-S3 (Seeed XIAO ESP32S3 Sense), an ESP32 (M5Stack ATOM) and Cortex-M33 parts (Arduino UNO Q, nRF5340, MCX N947).

conv1 1->8 3x3, ReLU, maxpool 2      28x28 -> 13x13
conv2 8->16 3x3, ReLU, maxpool 2     13x13 -> 5x5
fc1 400->32, ReLU
fc2 32->10

Input: (1, 1, 28, 28) float, pixel/255 normalised with mean 0.1307 and std 0.3081 (the usual MNIST constants; the firmware applies the same transform to camera crops). Output: 10 logits.

Files

File	What
tiny_digit_net.py	model definition (PyTorch)
tiny_digit_net.pt	trained state dict, fp32
executorch/fp32.pte	ExecuTorch program, portable fp32 kernels (62.8 KB)
executorch/int8-cmsis-m0plus.pte, -m4.pte, -m33.pte, -m7.pte	int8 via the ExecuTorch Cortex-M quantizer and CortexMPassManager, one per core (25.4 KB)
tflite/fp32.tflite, tflite/int8.tflite	the same weights through a Keras twin, for TensorFlow Lite Micro comparisons (post-training int8 quantization)
test_vectors.h	one test image plus reference logits for an on-device self-test

Results

Metric	Value
MNIST test accuracy (fp32)	97.46 %
Parameters	14,410
Inference, RP2040 @ 133 MHz, fp32 portable kernels	766 ms

Training

Standard MNIST, Adam, a few epochs on CPU; the script is train/train_mnist.py in the Pico4ML-BLE repo. Export with torch.export then to_edge_transform_and_lower(...).to_executorch(); the int8 variants use CortexMQuantizer + prepare_pt2e/convert_pt2e and the Cortex-M pass manager from ExecuTorch 1.5.

Where it runs

• executorch-zephyr-arducam-pico4ML-BLE: camera to screen digit reader
• executorch-zephyr-xiao-esp32s3: same on the XIAO ESP32-S3 Sense
• executorch-vs-tflm-benchmarks: ExecuTorch vs TFLM, fp32 and int8, across boards

License

Apache-2.0. MNIST is the public dataset by LeCun, Cortes and Burges.