Swin-Base: Optimized for Qualcomm Devices

SwinBase 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-Base 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-Base 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-Base 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: 88.8M
• Model size (float): 339 MB
• Model size (w8a16): 90.2 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Swin-Base	ONNX	float	Snapdragon® X2 Elite	8.384 ms	2 - 2 MB	NPU
Swin-Base	ONNX	float	Snapdragon® X Elite	19.605 ms	176 - 176 MB	NPU
Swin-Base	ONNX	float	Snapdragon® 8 Gen 3 Mobile	12.785 ms	0 - 536 MB	NPU
Swin-Base	ONNX	float	Snapdragon® 8 Gen 1 Mobile	29.317 ms	1 - 525 MB	NPU
Swin-Base	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	18.881 ms	1 - 4 MB	NPU
Swin-Base	ONNX	float	Qualcomm® QCS8450	29.317 ms	1 - 525 MB	NPU
Swin-Base	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	22.459 ms	0 - 4 MB	NPU
Swin-Base	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	19.605 ms	176 - 176 MB	NPU
Swin-Base	ONNX	float	Qualcomm® Dragonwing™ Q-8750	9.701 ms	1 - 387 MB	NPU
Swin-Base	ONNX	float	Snapdragon® 8 Elite Mobile	9.701 ms	1 - 387 MB	NPU
Swin-Base	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	7.959 ms	1 - 411 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® X2 Elite	6.967 ms	1 - 1 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® X Elite	16.251 ms	92 - 92 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	10.427 ms	0 - 538 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® 8 Gen 1 Mobile	20.499 ms	0 - 540 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® Dragonwing™ QCS6490	57.602 ms	0 - 3 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	15.409 ms	0 - 100 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® QCS8450	20.499 ms	0 - 540 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-9075	16.494 ms	0 - 3 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-X7181	16.251 ms	92 - 92 MB	NPU
Swin-Base	ONNX	w8a16	Qualcomm® Dragonwing™ Q-8750	8.176 ms	0 - 420 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® 8 Elite Mobile	8.176 ms	0 - 420 MB	NPU
Swin-Base	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	6.531 ms	0 - 434 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® X2 Elite	8.398 ms	1 - 1 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® X Elite	19.404 ms	1 - 1 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	12.423 ms	0 - 517 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	29.086 ms	0 - 502 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8275	53.644 ms	1 - 363 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	18.404 ms	1 - 3 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® SA8775P	21.38 ms	1 - 362 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® SA8650P	21.38 ms	1 - 362 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® SA8255P	21.38 ms	1 - 362 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® QCS8450	29.086 ms	0 - 502 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	19.404 ms	1 - 1 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	9.418 ms	1 - 366 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® SA7255P	53.644 ms	1 - 363 MB	NPU
Swin-Base	QNN_DLC	float	Qualcomm® SA8295P	27.587 ms	1 - 353 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® 8 Elite Mobile	9.418 ms	1 - 366 MB	NPU
Swin-Base	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	7.547 ms	0 - 386 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® X2 Elite	8.498 ms	0 - 0 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® X Elite	20.539 ms	0 - 0 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	12.791 ms	0 - 507 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8275	35.425 ms	0 - 412 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	19.158 ms	0 - 359 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® SA8775P	19.58 ms	0 - 411 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® SA8650P	19.58 ms	0 - 411 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® SA8255P	19.58 ms	0 - 411 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-9075	19.809 ms	0 - 2 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-X7181	20.539 ms	0 - 0 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-6690	123.141 ms	0 - 929 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-7790	21.594 ms	0 - 605 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-8750	9.514 ms	0 - 407 MB	NPU
Swin-Base	QNN_DLC	w8a16	Qualcomm® SA7255P	35.425 ms	0 - 412 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	9.514 ms	0 - 407 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	7.619 ms	0 - 421 MB	NPU
Swin-Base	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	21.594 ms	0 - 605 MB	NPU
Swin-Base	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	12.734 ms	0 - 1051 MB	NPU
Swin-Base	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	29.029 ms	0 - 511 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® Dragonwing™ QCS8275	53.464 ms	0 - 710 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	18.779 ms	0 - 4 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® SA8775P	21.969 ms	0 - 376 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® SA8650P	21.969 ms	0 - 376 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® SA8255P	21.969 ms	0 - 376 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® QCS8450	29.029 ms	0 - 511 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	22.083 ms	0 - 178 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	9.677 ms	0 - 379 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® SA7255P	53.464 ms	0 - 710 MB	NPU
Swin-Base	TFLITE	float	Qualcomm® SA8295P	28.357 ms	0 - 372 MB	NPU
Swin-Base	TFLITE	float	Snapdragon® 8 Elite Mobile	9.677 ms	0 - 379 MB	NPU
Swin-Base	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	7.935 ms	0 - 398 MB	NPU

License

• The license for the original implementation of Swin-Base 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.