Swin-Small: Optimized for Qualcomm Devices

SwinSmall is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This is based on the implementation of Swin-Small found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.45, ONNX Runtime 1.27.1	Download
ONNX	w8a16	Universal	QAIRT 2.45, ONNX Runtime 1.27.1	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
QNN_DLC	w8a16	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit Swin-Small on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for Swin-Small on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

• Model checkpoint: Imagenet
• Input resolution: 224x224
• Number of parameters: 50.4M
• Model size (float): 193 MB
• Model size (w8a16): 52.5 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Swin-Small	ONNX	float	Snapdragon® X2 Elite	6.803 ms	2 - 2 MB	NPU
Swin-Small	ONNX	float	Snapdragon® X Elite	16.29 ms	101 - 101 MB	NPU
Swin-Small	ONNX	float	Snapdragon® 8 Gen 3 Mobile	10.219 ms	0 - 419 MB	NPU
Swin-Small	ONNX	float	Snapdragon® 8 Gen 1 Mobile	23.899 ms	0 - 410 MB	NPU
Swin-Small	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	15.516 ms	0 - 8 MB	NPU
Swin-Small	ONNX	float	Qualcomm® QCS8450	23.899 ms	0 - 410 MB	NPU
Swin-Small	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	17.884 ms	0 - 4 MB	NPU
Swin-Small	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	16.29 ms	101 - 101 MB	NPU
Swin-Small	ONNX	float	Qualcomm® Dragonwing™ Q-8750	7.823 ms	1 - 606 MB	NPU
Swin-Small	ONNX	float	Snapdragon® 8 Elite Mobile	7.823 ms	1 - 606 MB	NPU
Swin-Small	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	6.424 ms	1 - 621 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® X2 Elite	5.98 ms	1 - 1 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® X Elite	13.798 ms	54 - 54 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	8.712 ms	0 - 528 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® 8 Gen 1 Mobile	16.407 ms	0 - 440 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® Dragonwing™ QCS6490	40.257 ms	0 - 3 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	13.191 ms	0 - 58 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® QCS8450	16.407 ms	0 - 440 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-9075	13.691 ms	0 - 3 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-X7181	13.798 ms	54 - 54 MB	NPU
Swin-Small	ONNX	w8a16	Qualcomm® Dragonwing™ Q-8750	6.926 ms	0 - 441 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® 8 Elite Mobile	6.926 ms	0 - 441 MB	NPU
Swin-Small	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	5.65 ms	0 - 453 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® X2 Elite	7.113 ms	1 - 1 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® X Elite	16.346 ms	1 - 1 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	9.934 ms	0 - 405 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	23.922 ms	0 - 396 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8275	37.942 ms	1 - 275 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	15.374 ms	1 - 4 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® SA8775P	17.388 ms	1 - 593 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® SA8650P	17.388 ms	1 - 593 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® SA8255P	17.388 ms	1 - 593 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® QCS8450	23.922 ms	0 - 396 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	17.609 ms	1 - 3 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	16.346 ms	1 - 1 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	7.617 ms	1 - 586 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® SA7255P	37.942 ms	1 - 275 MB	NPU
Swin-Small	QNN_DLC	float	Qualcomm® SA8295P	22.636 ms	1 - 269 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® 8 Elite Mobile	7.617 ms	1 - 586 MB	NPU
Swin-Small	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	6.248 ms	0 - 602 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® X2 Elite	7.213 ms	0 - 0 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® X Elite	17.313 ms	0 - 0 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	10.768 ms	0 - 870 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8275	29.258 ms	0 - 636 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	16.192 ms	0 - 3 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® SA8775P	16.763 ms	0 - 638 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® SA8650P	16.763 ms	0 - 638 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® SA8255P	16.763 ms	0 - 638 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-9075	17.063 ms	0 - 2 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-X7181	17.313 ms	0 - 0 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-6690	74.076 ms	0 - 752 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-7790	17.237 ms	0 - 652 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-8750	8.108 ms	0 - 622 MB	NPU
Swin-Small	QNN_DLC	w8a16	Qualcomm® SA7255P	29.258 ms	0 - 636 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	8.108 ms	0 - 622 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	6.502 ms	0 - 641 MB	NPU
Swin-Small	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	17.237 ms	0 - 652 MB	NPU
Swin-Small	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	10.204 ms	0 - 433 MB	NPU
Swin-Small	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	23.391 ms	0 - 410 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® Dragonwing™ QCS8275	37.439 ms	0 - 302 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	15.556 ms	0 - 4 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® SA8775P	17.487 ms	0 - 306 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® SA8650P	17.487 ms	0 - 306 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® SA8255P	17.487 ms	0 - 306 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® QCS8450	23.391 ms	0 - 410 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	17.613 ms	0 - 104 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	7.825 ms	0 - 303 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® SA7255P	37.439 ms	0 - 302 MB	NPU
Swin-Small	TFLITE	float	Qualcomm® SA8295P	22.998 ms	0 - 297 MB	NPU
Swin-Small	TFLITE	float	Snapdragon® 8 Elite Mobile	7.825 ms	0 - 303 MB	NPU
Swin-Small	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	6.601 ms	0 - 322 MB	NPU

License

• The license for the original implementation of Swin-Small can be found here.

References

• Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.