Overview of PyTorch STM32 Model Zoo for Image Classification

The STM32 model zoo includes several PyTorch-based models for image classification use cases, converted to ONNX format and optimized for STM32N6 NPU deployment. All models are pre-trained on ImageNet and and converted to ONNX format with QDQ quantization.

Model Categories

• Public_pretrainedmodel_public_dataset contains public PyTorch image classification models trained on public datasets and converted to ONNX format with QDQ quantization.
• ST_pretrainedmodel_public_dataset contains public PyTorch image classification models trained on public dataset by ST.

Explore the complete PyTorch model zoo with pre-trained models optimized for STM32N6.

Model Families

You can get comprehensive footprints and performance information for each model family following the links below:

Mobile-Optimized Architectures

• mobilenet_pt – Efficient depthwise separable convolutions (alpha variants: 0.25, 0.50, 0.75)
• mobilenetv2_pt – Inverted residual blocks with linear bottlenecks (alpha/width variants)
• mobilenetv4_pt – Latest MobileNet iteration with enhanced efficiency
• fdmobilenet_pt – Fast downsampling variants for reduced latency

Lightweight Efficient Models

• shufflenetv2_pt – Channel shuffle operations (0.5x, 1.0x width multipliers)
• mnasnet_pt – Mobile neural architecture search optimized models
• regnet_pt – Network design space exploration

STMicroelectronics In-house Model

• st_resnet_pt – Custom ResNet variants optimized for STM32 (pico, nano, micro, milli, tiny)

Standard Architectures

• resnet18_pt – Classic residual networks (width variant 0.25)
• preresnet18_pt – Pre-activation ResNet variant
• dla_pt – Deep Layer Aggregation with hierarchical features
• squeezenet_pt – Fire modules with squeeze and expand
• sqnxt_pt – More efficient SqueezeNet variants

Specialized Architectures

• hardnet_pt – Harmonic DenseNet design
• peleenet_pt – Efficient DenseNet variant
• proxylessnas_pt – Direct NAS on target hardware
• semnasnet_pt – Squeeze-and-Excitation MnasNet
• darknet_pt – Lightweight feature extraction backbone

Quick Selection Guide

By Inference Speed (on STM32N6570-DK)

• Ultra-Fast (<5ms): fdmobilenet_a025 (1.88ms), mobilenet_a025 (2.98ms)
• Fast (5-10ms): fdmobilenet_a050 (4.07ms), mobilenet_a050 (6.55ms), shufflenetv2_x050 (8.35ms)
• Moderate (10-20ms): mobilenetv2_a050 (10.08ms), mobilenetv4small (13.74ms), st_resnetmicro (13.83ms)
• Balanced (20-40ms): mobilenetv2_a100 (20.35ms), st_resnettiny (24.10ms), proxylessnas (27.65ms)

By Model Size

• Tiny (<500KB): fdmobilenet_a025 (377KB), mobilenet_a025 (469KB), st_resnetpico (607KB)
• Small (500KB-1.5MB): fdmobilenet_a050 (973KB), mobilenet_a050 (1.3MB), shufflenetv2_x050 (1.4MB)
• Medium (1.5-3MB): mobilenetv2_a050 (1.97MB), mnasnet_d050 (2.3MB), peleenet (2.75MB)
• Large (>3MB): mobilenetv4small (3.76MB), mobilenetv2_a100 (3.81MB), st_resnettiny (4.06MB)

By RAM Requirements

• Internal RAM Only (<1MB): Most MobileNet, ShuffleNet, and ST ResNet variants
• Requires External RAM: DLA family (~6MB), SqueezeNext family (3-9MB), SqueezeNetv10 (~6.7MB)

Platform Support

All PyTorch models are specifically optimized for:

• Primary Target: STM32N6570-DK with NPU acceleration
• Framework: ONNX with QDQ INT8 quantization
• STEdgeAI Core: Version 3.0.0+

Performance Notes

• All inference times measured on STM32N6570-DK with NPU/MCU execution
• Internal RAM usage varies from 294 KiB (fdmobilenet_a025) to 11,350 KiB (sqnxt23_x200)
• Models using only internal RAM are preferred for simpler deployment
• External RAM models (DLA, SqueezeNext) offer different accuracy/performance tradeoffs

Feel free to explore the model zoo and get pre-trained models here.

For training and deployment guidance, refer to the STM32 AI model zoo documentation.