Use case : Image classification

Model description

ProxylessNAS enables direct neural architecture search on target hardware, eliminating the "proxy" task typically used in NAS. It learns specialized architectures optimized for specific devices without costly re-training.

The architecture employs direct hardware targeting by searching directly on target hardware metrics, using path-level binarization as an efficient search method with binary architecture parameters. Latency regularization incorporates actual latency into the search objective, resulting in hardware-specific architectures optimized for different hardware platforms.

ProxylessNAS achieves high accuracy (74.25% Top-1) with good quantization stability (0.60% drop), making it ideal for applications requiring hardware-optimized architectures with strict latency requirements.

(source: https://arxiv.org/abs/1812.00332)

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

Network information

Network Information	Value
Framework	Torch
MParams	~4.13 M
Quantization	Int8
Provenance	https://github.com/mit-han-lab/proxylessnas
Paper	https://arxiv.org/abs/1812.00332

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
proxylessnas_pt_224	Imagenet	Int8	224×224×3	STM32N6	1372	0	4233.20	4.0.0

Reference NPU inference time on food101 and imagenet dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STEdgeAI Core version
proxylessnas_pt_224	Imagenet	Int8	224×224×3	STM32N6570-DK	NPU/MCU	27.65	36.17	4.0.0

Accuracy with Imagenet dataset

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
proxylessnas_pt	Float	224x224x3	74.85 %
proxylessnas_pt	Int8	224x224x3	74.25 %

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: ProxylessNAS — https://github.com/MIT-HAN-LAB/ProxylessNAS