STMicroelectronics/st_resnet_pt

ST ResNet

Use case : Image classification

Model description

ST ResNet is STMicroelectronics' custom ResNet family specifically designed and optimized for STM32 deployment. It offers a range of sizes from "pico" to "tiny" with ReLU activations, providing a progressive accuracy-efficiency trade-off tailored for embedded vision applications.

The architecture is STM32-optimized and designed specifically for STM32 NPU deployment, with progressive sizing from Pico → Nano → Micro → Milli → Tiny (increasing capacity). It uses ReLU activation for quantization friendliness and a balanced design optimized for both accuracy and inference speed on target hardware.

ST ResNet models are the recommended choice for production STM32 deployments, with all variants running on internal RAM only and well-characterized performance on STM32 hardware.

(source: https://arxiv.org/abs/2601.05364, https://arxiv.org/abs/2511.11716)

The model is quantized to int8 using ONNX Runtime and exported for efficient deployment.

Network information

Network Information	Value
Framework	Torch
MParams	~0.59–3.97 M
Quantization	Int8
Provenance	https://github.com/STMicroelectronics/stm32ai-modelzoo
Paper	https://arxiv.org/abs/2601.05364

Network inputs / outputs

For an image resolution of NxM and P classes

Input Shape	Description
(1, N, M, 3)	Single NxM RGB image with UINT8 values between 0 and 255

Output Shape	Description
(1, P)	Per-class confidence for P classes in FLOAT32

Recommended platforms

Platform	Supported	Recommended
STM32L0	[]	[]
STM32L4	[]	[]
STM32U5	[]	[]
STM32H7	[]	[]
STM32MP1	[]	[]
STM32MP2	[]	[]
STM32N6	[x]	[x]

Performances

Metrics

• Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
• All the models are trained from scratch on Imagenet dataset

Reference NPU memory footprint on Imagenet dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Series	Internal RAM (KiB)	External RAM (KiB)	Weights Flash (KiB)	STEdgeAI Core version
st_resnetpico_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6	784	0	607.27	3.0.0
st_resnetnano_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6	759.5	0	992.04	3.0.0
st_resnetmicro_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6	882	0	1534.12	3.0.0
st_resnetmilli_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6	1421	0	3059.81	3.0.0
st_resnettiny_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6	2205	0	4060.57	3.0.0

Reference NPU inference time on imagenet dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STEdgeAI Core version
st_resnetpico_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	7.54	132.63	3.0.0
st_resnetnano_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	9.46	105.71	3.0.0
st_resnetmicro_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	13.83	72.31	3.0.0
st_resnetmilli_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	17.59	56.85	3.0.0
st_resnettiny_actrelu_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	24.10	41.49	3.0.0

Model	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STM32Cube.AI version	STEdgeAI Core version
st_resnetpico_actrelu_pt_224	Int8	224×224×3	STM32N6570-DK	NPU/MCU	7.54	132.63	10.2.0	2.2.0
st_resnetnano_actrelu_pt_224	Int8	224×224×3	STM32N6570-DK	NPU/MCU	9.46	105.71	10.2.0	2.2.0
st_resnetmicro_actrelu_pt_224	Int8	224×224×3	STM32N6570-DK	NPU/MCU	13.83	72.31	10.2.0	2.2.0
st_resnetmilli_actrelu_pt_224	Int8	224×224×3	STM32N6570-DK	NPU/MCU	17.59	56.85	10.2.0	2.2.0
st_resnettiny_actrelu_pt_224	Int8	224×224×3	STM32N6570-DK	NPU/MCU	24.10	41.49	10.2.0	2.2.0

Accuracy with Imagenet dataset

Model	Format	Resolution	Top 1 Accuracy
st_resnetmicro_actrelu_pt	Float	224×224×3	66.43 %
st_resnetmicro_actrelu_pt	Int8	224×224×3	65.62 %
st_resnetmilli_actrelu_pt	Float	224×224×3	71.10 %
st_resnetmilli_actrelu_pt	Int8	224×224×3	70.45 %
st_resnetnano_actrelu_pt	Float	224×224×3	59.32 %
st_resnetnano_actrelu_pt	Int8	224×224×3	58.25 %
st_resnetpico_actrelu_pt	Float	224×224×3	49.42 %
st_resnetpico_actrelu_pt	Int8	224×224×3	46.98 %
st_resnettiny_actrelu_pt	Float	224×224×3	72.07 %
st_resnettiny_actrelu_pt	Int8	224×224×3	71.40 %

Dataset details: link Number of classes: 1000. To perform the quantization, we calibrated the activations with a random subset of the training set. For the sake of simplicity, the accuracy reported here was estimated on the 50000 labelled images of the validation set.

Model	Format	Resolution	Top 1 Accuracy
st_resnetmicro_actrelu_pt	Float	224×224×3	66.43 %
st_resnetmicro_actrelu_pt	Int8	224×224×3	65.62 %
st_resnetmilli_actrelu_pt	Float	224×224×3	71.10 %
st_resnetmilli_actrelu_pt	Int8	224×224×3	70.45 %
st_resnetnano_actrelu_pt	Float	224×224×3	59.32 %
st_resnetnano_actrelu_pt	Int8	224×224×3	58.25 %
st_resnetpico_actrelu_pt	Float	224×224×3	49.42 %
st_resnetpico_actrelu_pt	Int8	224×224×3	46.98 %
st_resnettiny_actrelu_pt	Float	224×224×3	72.07 %
st_resnettiny_actrelu_pt	Int8	224×224×3	71.40 %

Retraining and Integration in a simple example:

Please refer to the stm32ai-modelzoo-services GitHub here

References

[1] - Dataset: Imagenet (ILSVRC 2012) — https://www.image-net.org/

[2] - Model: STRESNET & STYOLO — https://arxiv.org/abs/2601.05364

[3] - Model: CompressNAS — https://arxiv.org/abs/2511.11716