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 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a16	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a16	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 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® X Elite	587.026 ms	187 - 187 MB	NPU
Video-MAE	ONNX	float	Snapdragon® 8 Gen 3 Mobile	368.95 ms	2 - 1260 MB	NPU
Video-MAE	ONNX	float	Qualcomm® QCS8550 (Proxy)	568.266 ms	0 - 5 MB	NPU
Video-MAE	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	390.549 ms	2 - 1045 MB	NPU
Video-MAE	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	459.094 ms	9 - 1089 MB	NPU
Video-MAE	ONNX	float	Snapdragon® X2 Elite	448.083 ms	187 - 187 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® X Elite	472.261 ms	9 - 9 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	381.871 ms	9 - 1170 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	1091.288 ms	1 - 948 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	453.084 ms	9 - 12 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8775P	497.143 ms	0 - 972 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® QCS9075	514.195 ms	9 - 20 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	580.221 ms	9 - 1067 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA7255P	1091.288 ms	1 - 948 MB	NPU
Video-MAE	QNN_DLC	float	Qualcomm® SA8295P	569.0 ms	0 - 858 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	260.228 ms	9 - 969 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	329.664 ms	9 - 962 MB	NPU
Video-MAE	QNN_DLC	float	Snapdragon® X2 Elite	295.799 ms	9 - 9 MB	NPU
Video-MAE	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	101.855 ms	0 - 1168 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® QCS8275 (Proxy)	5585.078 ms	42 - 59 MB	CPU
Video-MAE	TFLITE	float	Qualcomm® QCS8550 (Proxy)	140.998 ms	0 - 4 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA8775P	161.076 ms	0 - 955 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® QCS9075	172.2 ms	0 - 207 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® QCS8450 (Proxy)	296.47 ms	1 - 1113 MB	NPU
Video-MAE	TFLITE	float	Qualcomm® SA7255P	5585.078 ms	42 - 59 MB	CPU
Video-MAE	TFLITE	float	Qualcomm® SA8295P	214.768 ms	0 - 907 MB	NPU
Video-MAE	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	74.046 ms	0 - 965 MB	NPU
Video-MAE	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	60.611 ms	0 - 965 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.