SEMnasNet

Use case : Image classification

Model description

SEMnasNet combines the MnasNet architecture with Squeeze-and-Excitation (SE) blocks, adding channel attention mechanisms to the NAS-derived architecture for improved accuracy.

The architecture builds on MnasNet's NAS-derived efficient design and adds Squeeze-and-Excitation blocks for channel attention and feature recalibration. Adaptive feature weighting emphasizes informative channels, with SE blocks boosting accuracy with minimal overhead.

SEMnasNet achieves the highest accuracy in the model zoo (75.38% Top-1) with excellent quantization stability (0.37% drop), making it the best choice for accuracy-critical applications.

(source: https://arxiv.org/abs/1807.11626, https://arxiv.org/abs/1709.01507)

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

Network information

Network inputs / outputs

For an image resolution of NxM and P classes

Recommended platforms

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)

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

Accuracy with Imagenet dataset

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 (MnasNet): MnasNet — https://arxiv.org/abs/1807.11626

[3] - Model (SE-Net): Squeeze-and-Excitation Networks — https://arxiv.org/abs/1709.01507