Video-MAE: Optimized for Qualcomm Devices

Video MAE (Masked Auto Encoder) is a network for doing video classification that uses the ViT (Vision Transformer) backbone.

This is based on the implementation of Video-MAE 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
QNN_DLC	float	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

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

Model Details

Model Type: Model_use_case.video_classification

Model Stats:

• Model checkpoint: Kinectics-400
• Input resolution: 224x224
• Number of parameters: 87.7M
• Model size (float): 335 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
Video-MAE	ONNX	float	Snapdragon® X2 Elite	1580.262 ms	46 - 46 MB	NPU
Video-MAE	ONNX	float	Snapdragon® X Elite	2859.465 ms	192 - 192 MB	NPU
Video-MAE	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	2838.231 ms	1 - 213 MB	NPU
Video-MAE	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	4901.136 ms	46 - 95 MB	NPU
Video-MAE	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	2859.465 ms	192 - 192 MB	NPU
Video-MAE	ONNX	float	Qualcomm® Dragonwing™ Q-8750	1643.62 ms	1 - 5495 MB	NPU
Video-MAE	ONNX	float	Snapdragon® 8 Elite Mobile	1643.62 ms	1 - 5495 MB	NPU
Video-MAE	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	1515.805 ms	0 - 5707 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® X2 Elite	1967.524 ms	46 - 46 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® X Elite	3232.14 ms	46 - 46 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	5141.062 ms	46 - 49 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8775P	5350.709 ms	36 - 6136 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8650P	5350.709 ms	36 - 6136 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8255P	5350.709 ms	36 - 6136 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	5288.941 ms	46 - 94 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	3232.14 ms	46 - 46 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	3552.153 ms	1 - 6092 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8295P	3815.178 ms	36 - 5739 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® 8 Elite Mobile	3552.153 ms	1 - 6092 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	4073.389 ms	7 - 6274 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	5170.295 ms	1 - 5 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA8775P	5319.772 ms	2 - 5973 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA8650P	5319.772 ms	2 - 5973 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA8255P	5319.772 ms	2 - 5973 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	5259.479 ms	0 - 279 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	3540.118 ms	3 - 5954 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA8295P	3741.165 ms	2 - 5604 MB	NPU
Video-MAE	TFLITE	float	Snapdragon® 8 Elite Mobile	3540.118 ms	3 - 5954 MB	NPU
Video-MAE	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	4085.554 ms	1 - 6105 MB	NPU

License

• The license for the original implementation of Video-MAE can be found here.

References

• Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
• 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.