ResNet-2Plus1D: Optimized for Qualcomm Devices

ResNet (2+1)D Convolutions is a network which explicitly factorizes 3D convolution into two separate and successive operations, a 2D spatial convolution and a 1D temporal convolution. It used for video understanding applications.

This is based on the implementation of ResNet-2Plus1D 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
TFLITE	w8a8	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit ResNet-2Plus1D 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 ResNet-2Plus1D on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.video_classification

Model Stats:

• Input resolution: 112x112
• Model checkpoint: Kinetics-400
• Model size (float): 120 MB
• Model size (w8a8): 30.8 MB
• Number of parameters: 31.5M

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
ResNet-2Plus1D	ONNX	float	Snapdragon® X2 Elite	6.167 ms	2 - 2 MB	NPU
ResNet-2Plus1D	ONNX	float	Snapdragon® X Elite	12.131 ms	60 - 60 MB	NPU
ResNet-2Plus1D	ONNX	float	Snapdragon® 8 Gen 3 Mobile	9.077 ms	1 - 279 MB	NPU
ResNet-2Plus1D	ONNX	float	Snapdragon® 8 Gen 1 Mobile	25.811 ms	1 - 260 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® Dragonwing™ IQ-8275	24.088 ms	2 - 8 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	12.011 ms	0 - 78 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® QCS8450	25.811 ms	1 - 260 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	21.247 ms	2 - 7 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	12.131 ms	60 - 60 MB	NPU
ResNet-2Plus1D	ONNX	float	Qualcomm® Dragonwing™ Q-8750	7.174 ms	0 - 206 MB	NPU
ResNet-2Plus1D	ONNX	float	Snapdragon® 8 Elite Mobile	7.174 ms	0 - 206 MB	NPU
ResNet-2Plus1D	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	5.614 ms	1 - 211 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® X2 Elite	6.642 ms	2 - 2 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® X Elite	12.844 ms	2 - 2 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	9.174 ms	0 - 289 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	28.462 ms	2 - 273 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	23.819 ms	2 - 7 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-8275	81.653 ms	1 - 216 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	12.431 ms	2 - 4 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® SA8775P	21.217 ms	0 - 216 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® SA8650P	21.217 ms	0 - 216 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® SA8255P	21.217 ms	0 - 216 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® QCS8450	28.462 ms	2 - 273 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	22.161 ms	4 - 8 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	12.844 ms	2 - 2 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	7.193 ms	2 - 222 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® SA7255P	81.653 ms	1 - 216 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Qualcomm® SA8295P	22.771 ms	0 - 198 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® 8 Elite Mobile	7.193 ms	2 - 222 MB	NPU
ResNet-2Plus1D	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	5.739 ms	2 - 229 MB	NPU
ResNet-2Plus1D	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	290.49 ms	0 - 316 MB	NPU
ResNet-2Plus1D	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	437.002 ms	1 - 301 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	365.131 ms	0 - 67 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® Dragonwing™ IQ-8275	697.289 ms	0 - 238 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	374.944 ms	0 - 2 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® SA8775P	378.003 ms	0 - 237 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® SA8650P	378.003 ms	0 - 237 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® SA8255P	378.003 ms	0 - 237 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® QCS8450	437.002 ms	1 - 301 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	393.467 ms	0 - 66 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	266.832 ms	0 - 241 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® SA7255P	697.289 ms	0 - 238 MB	NPU
ResNet-2Plus1D	TFLITE	float	Qualcomm® SA8295P	459.982 ms	0 - 231 MB	NPU
ResNet-2Plus1D	TFLITE	float	Snapdragon® 8 Elite Mobile	266.832 ms	0 - 241 MB	NPU
ResNet-2Plus1D	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	255.1 ms	0 - 250 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	576.776 ms	0 - 516 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Snapdragon® 8 Gen 1 Mobile	867.653 ms	1 - 472 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ QCS6490	1784.014 ms	267 - 428 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ IQ-8275	760.312 ms	0 - 67 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ IQ-8275	1520.639 ms	1 - 439 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ QCS8550 (Proxy)	751.66 ms	0 - 4 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® SA8775P	799.676 ms	1 - 441 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® SA8650P	799.676 ms	1 - 441 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® SA8255P	799.676 ms	1 - 441 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® QCS8450	867.653 ms	1 - 472 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ IQ-9075	797.865 ms	0 - 66 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ Q-6690	1613.816 ms	290 - 451 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ Q-7790	1228.272 ms	252 - 324 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® Dragonwing™ Q-8750	508.363 ms	0 - 457 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® SA7255P	1520.639 ms	1 - 439 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Qualcomm® SA8295P	864.863 ms	1 - 436 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Snapdragon® 8 Elite Mobile	508.363 ms	0 - 457 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	639.353 ms	0 - 480 MB	NPU
ResNet-2Plus1D	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	1228.272 ms	252 - 324 MB	NPU

License

• The license for the original implementation of ResNet-2Plus1D can be found here.

References

• A Closer Look at Spatiotemporal Convolutions for Action 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.