Swin-Tiny: Optimized for Qualcomm Devices

SwinTiny 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-Tiny 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-Tiny 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-Tiny 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: 28.8M
• Model size (float): 110 MB
• Model size (w8a16): 29.9 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Swin-Tiny	ONNX	float	Snapdragon® X2 Elite	3.98 ms	2 - 2 MB	NPU
Swin-Tiny	ONNX	float	Snapdragon® X Elite	9.584 ms	57 - 57 MB	NPU
Swin-Tiny	ONNX	float	Snapdragon® 8 Gen 3 Mobile	5.974 ms	0 - 253 MB	NPU
Swin-Tiny	ONNX	float	Snapdragon® 8 Gen 1 Mobile	14.172 ms	0 - 245 MB	NPU
Swin-Tiny	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	9.101 ms	1 - 4 MB	NPU
Swin-Tiny	ONNX	float	Qualcomm® QCS8450	14.172 ms	0 - 245 MB	NPU
Swin-Tiny	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	10.193 ms	0 - 4 MB	NPU
Swin-Tiny	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	9.584 ms	57 - 57 MB	NPU
Swin-Tiny	ONNX	float	Qualcomm® Dragonwing™ Q-8750	4.553 ms	0 - 434 MB	NPU
Swin-Tiny	ONNX	float	Snapdragon® 8 Elite Mobile	4.553 ms	0 - 434 MB	NPU
Swin-Tiny	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	3.78 ms	0 - 433 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® X2 Elite	3.464 ms	1 - 1 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® X Elite	8.224 ms	31 - 31 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® 8 Gen 3 Mobile	5.161 ms	0 - 335 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® 8 Gen 1 Mobile	9.71 ms	0 - 278 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® Dragonwing™ QCS6490	25.234 ms	0 - 3 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	7.797 ms	0 - 36 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® QCS8450	9.71 ms	0 - 278 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-9075	8.122 ms	0 - 3 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® Dragonwing™ IQ-X7181	8.224 ms	31 - 31 MB	NPU
Swin-Tiny	ONNX	w8a16	Qualcomm® Dragonwing™ Q-8750	3.967 ms	0 - 266 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® 8 Elite Mobile	3.967 ms	0 - 266 MB	NPU
Swin-Tiny	ONNX	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	3.269 ms	0 - 276 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® X2 Elite	4.268 ms	1 - 1 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® X Elite	9.9 ms	1 - 1 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	5.963 ms	0 - 248 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	14.523 ms	0 - 240 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8275	21.24 ms	1 - 175 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	9.102 ms	1 - 4 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® SA8775P	10.419 ms	1 - 435 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® SA8650P	10.419 ms	1 - 435 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® SA8255P	10.419 ms	1 - 435 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® QCS8450	14.523 ms	0 - 240 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	10.252 ms	1 - 3 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	9.9 ms	1 - 1 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	4.514 ms	0 - 414 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® SA7255P	21.24 ms	1 - 175 MB	NPU
Swin-Tiny	QNN_DLC	float	Qualcomm® SA8295P	13.806 ms	0 - 433 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® 8 Elite Mobile	4.514 ms	0 - 414 MB	NPU
Swin-Tiny	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	3.772 ms	0 - 412 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® X2 Elite	4.462 ms	0 - 0 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® X Elite	10.929 ms	0 - 0 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® 8 Gen 3 Mobile	6.605 ms	0 - 508 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8275	17.796 ms	0 - 437 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ QCS8550 (Proxy)	10.199 ms	0 - 3 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® SA8775P	10.647 ms	0 - 442 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® SA8650P	10.647 ms	0 - 442 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® SA8255P	10.647 ms	0 - 442 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-9075	10.49 ms	2 - 4 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ IQ-X7181	10.929 ms	0 - 0 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-6690	44.109 ms	0 - 430 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-7790	10.579 ms	0 - 435 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® Dragonwing™ Q-8750	4.883 ms	0 - 431 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Qualcomm® SA7255P	17.796 ms	0 - 437 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® 8 Elite Mobile	4.883 ms	0 - 431 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® 8 Elite Gen 5 Mobile	3.893 ms	0 - 440 MB	NPU
Swin-Tiny	QNN_DLC	w8a16	Snapdragon® 7 Gen 4 Mobile	10.579 ms	0 - 435 MB	NPU
Swin-Tiny	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	6.028 ms	34 - 300 MB	NPU
Swin-Tiny	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	14.341 ms	0 - 259 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® Dragonwing™ QCS8275	21.22 ms	0 - 200 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	9.247 ms	0 - 3 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® SA8775P	10.427 ms	0 - 201 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® SA8650P	10.427 ms	0 - 201 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® SA8255P	10.427 ms	0 - 201 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® QCS8450	14.341 ms	0 - 259 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	10.18 ms	0 - 60 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	4.632 ms	0 - 205 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® SA7255P	21.22 ms	0 - 200 MB	NPU
Swin-Tiny	TFLITE	float	Qualcomm® SA8295P	13.875 ms	0 - 199 MB	NPU
Swin-Tiny	TFLITE	float	Snapdragon® 8 Elite Mobile	4.632 ms	0 - 205 MB	NPU
Swin-Tiny	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	3.837 ms	0 - 214 MB	NPU

License

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