EfficientNetV2: Smaller Models and Faster Training
Paper • 2104.00298 • Published • 1
EfficientNet-V2-s: Optimized for Qualcomm Devices
EfficientNetV2-s 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 EfficientNet-V2-s 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 EfficientNet-V2-s 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 EfficientNet-V2-s on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.image_classification
Model Stats:
- Input resolution: 384x384
- Model checkpoint: Imagenet
- Model size (float): 81.7 MB
- Model size (w8a16): 27.2 MB
- Number of parameters: 21.4M
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| EfficientNet-V2-s | ONNX | float | Snapdragon® X2 Elite | 3.041 ms | 2 - 2 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Snapdragon® X Elite | 5.591 ms | 46 - 46 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 4.055 ms | 0 - 165 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 11.642 ms | 1 - 197 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 8.114 ms | 2 - 7 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 5.353 ms | 0 - 232 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® QCS8450 | 11.642 ms | 1 - 197 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 7.648 ms | 1 - 6 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 5.591 ms | 46 - 46 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 3.1 ms | 0 - 202 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Snapdragon® 8 Elite Mobile | 3.1 ms | 0 - 202 MB | NPU |
| EfficientNet-V2-s | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.318 ms | 0 - 205 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® X2 Elite | 2.363 ms | 2 - 2 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® X Elite | 5.696 ms | 24 - 24 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 3.641 ms | 0 - 207 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® 8 Gen 1 Mobile | 6.733 ms | 1 - 215 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS6490 | 20.194 ms | 1 - 4 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 5.131 ms | 1 - 4 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 5.453 ms | 0 - 240 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® QCS8450 | 6.733 ms | 1 - 215 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 5.788 ms | 1 - 4 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 5.696 ms | 24 - 24 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 39.079 ms | 0 - 271 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 6.572 ms | 1 - 271 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 2.5 ms | 0 - 162 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® 8 Elite Mobile | 2.5 ms | 0 - 162 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 2.021 ms | 0 - 165 MB | NPU |
| EfficientNet-V2-s | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 6.572 ms | 1 - 271 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® X2 Elite | 3.435 ms | 2 - 2 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® X Elite | 6.384 ms | 2 - 2 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 4.343 ms | 0 - 158 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 13.548 ms | 0 - 188 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 8.213 ms | 2 - 6 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 5.804 ms | 2 - 4 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® SA8775P | 8.209 ms | 2 - 76 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® SA8650P | 8.209 ms | 2 - 76 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® SA8255P | 8.209 ms | 2 - 76 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® QCS8450 | 13.548 ms | 0 - 188 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 7.909 ms | 2 - 5 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 6.384 ms | 2 - 2 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 3.132 ms | 0 - 78 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® SA7255P | 25.729 ms | 2 - 75 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Qualcomm® SA8295P | 13.03 ms | 2 - 107 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 3.132 ms | 0 - 78 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.36 ms | 2 - 85 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® X2 Elite | 2.913 ms | 1 - 1 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® X Elite | 6.767 ms | 1 - 1 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 4.183 ms | 0 - 175 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® 8 Gen 1 Mobile | 8.035 ms | 1 - 187 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS6490 | 21.432 ms | 1 - 3 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-8275 | 5.795 ms | 1 - 4 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 6.242 ms | 1 - 3 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® SA8775P | 6.716 ms | 1 - 136 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® SA8650P | 6.716 ms | 1 - 136 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® SA8255P | 6.716 ms | 1 - 136 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® QCS8450 | 8.035 ms | 1 - 187 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-9075 | 6.473 ms | 3 - 5 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ IQ-X7181 | 6.767 ms | 1 - 1 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-6690 | 40.733 ms | 1 - 253 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-7790 | 7.206 ms | 1 - 254 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® Dragonwing™ Q-8750 | 2.825 ms | 0 - 143 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® SA7255P | 12.076 ms | 1 - 134 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Qualcomm® SA8295P | 8.38 ms | 1 - 134 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® 8 Elite Mobile | 2.825 ms | 0 - 143 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 2.245 ms | 1 - 146 MB | NPU |
| EfficientNet-V2-s | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 7.206 ms | 1 - 254 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 4.324 ms | 0 - 201 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 13.403 ms | 0 - 230 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 8.214 ms | 0 - 52 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 5.756 ms | 0 - 2 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® SA8775P | 8.252 ms | 0 - 118 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® SA8650P | 8.252 ms | 0 - 118 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® SA8255P | 8.252 ms | 0 - 118 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® QCS8450 | 13.403 ms | 0 - 230 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 7.951 ms | 0 - 51 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 3.135 ms | 0 - 124 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® SA7255P | 25.847 ms | 0 - 115 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Qualcomm® SA8295P | 13.025 ms | 0 - 148 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Snapdragon® 8 Elite Mobile | 3.135 ms | 0 - 124 MB | NPU |
| EfficientNet-V2-s | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 2.33 ms | 0 - 120 MB | NPU |
License
- The license for the original implementation of EfficientNet-V2-s can be found here.
References
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.