NASNet: Optimized for Qualcomm Devices

NASNet is a vision transformer model that can classify images from the Imagenet dataset.

This is based on the implementation of NASNet 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 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a8_mixed_fp16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a8_mixed_fp16	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit NASNet 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 NASNet on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

• Model checkpoint: nasnetalarge.tf_in1k
• Input resolution: 224x224
• GMACs: 5.9
• Activations (M): 19.4
• Number of parameters: 88.7M
• Model size (float): 338 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
NASNet	ONNX	float	Snapdragon® X2 Elite	8.215 ms	189 - 189 MB	NPU
NASNet	ONNX	float	Snapdragon® X Elite	17.826 ms	188 - 188 MB	NPU
NASNet	ONNX	float	Snapdragon® 8 Gen 3 Mobile	12.748 ms	0 - 831 MB	NPU
NASNet	ONNX	float	Qualcomm® QCS8550 (Proxy)	17.461 ms	0 - 196 MB	NPU
NASNet	ONNX	float	Qualcomm® QCS9075	28.336 ms	0 - 4 MB	NPU
NASNet	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	10.601 ms	1 - 676 MB	NPU
NASNet	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	8.389 ms	1 - 673 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Snapdragon® X2 Elite	4.546 ms	100 - 100 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Snapdragon® X Elite	11.742 ms	98 - 98 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Snapdragon® 8 Gen 3 Mobile	6.989 ms	5 - 481 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Qualcomm® QCS8550 (Proxy)	9.916 ms	5 - 11 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Qualcomm® QCS9075	11.712 ms	5 - 8 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Snapdragon® 8 Elite For Galaxy Mobile	5.62 ms	0 - 346 MB	NPU
NASNet	ONNX	w8a8_mixed_fp16	Snapdragon® 8 Elite Gen 5 Mobile	4.568 ms	5 - 359 MB	NPU
NASNet	QNN_DLC	float	Snapdragon® X2 Elite	9.24 ms	1 - 1 MB	NPU
NASNet	QNN_DLC	float	Snapdragon® X Elite	19.234 ms	1 - 1 MB	NPU
NASNet	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	12.326 ms	0 - 814 MB	NPU
NASNet	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	54.016 ms	1 - 659 MB	NPU
NASNet	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	19.0 ms	1 - 3 MB	NPU
NASNet	QNN_DLC	float	Qualcomm® QCS9075	28.724 ms	1 - 3 MB	NPU
NASNet	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	35.267 ms	0 - 792 MB	NPU
NASNet	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	11.072 ms	1 - 652 MB	NPU
NASNet	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	8.904 ms	1 - 654 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Snapdragon® X2 Elite	4.134 ms	0 - 0 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Snapdragon® X Elite	9.236 ms	0 - 0 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Snapdragon® 8 Gen 3 Mobile	5.992 ms	0 - 495 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Qualcomm® QCS8275 (Proxy)	16.472 ms	0 - 378 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Qualcomm® QCS8550 (Proxy)	8.936 ms	0 - 2 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Qualcomm® QCS9075	9.454 ms	0 - 2 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Qualcomm® QCS8450 (Proxy)	11.194 ms	0 - 507 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Snapdragon® 8 Elite For Galaxy Mobile	4.926 ms	0 - 377 MB	NPU
NASNet	QNN_DLC	w8a8_mixed_fp16	Snapdragon® 8 Elite Gen 5 Mobile	3.895 ms	0 - 378 MB	NPU
NASNet	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	8.796 ms	0 - 783 MB	NPU
NASNet	TFLITE	float	Qualcomm® QCS8275 (Proxy)	44.427 ms	0 - 628 MB	NPU
NASNet	TFLITE	float	Qualcomm® QCS8550 (Proxy)	12.456 ms	0 - 3 MB	NPU
NASNet	TFLITE	float	Qualcomm® QCS9075	15.469 ms	0 - 192 MB	NPU
NASNet	TFLITE	float	Qualcomm® QCS8450 (Proxy)	28.855 ms	0 - 762 MB	NPU
NASNet	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	6.956 ms	0 - 623 MB	NPU
NASNet	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	5.632 ms	0 - 625 MB	NPU

License

• The license for the original implementation of NASNet can be found here.

References

• Learning Transferable Architectures for Scalable Image Recognition
• 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.