MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer
Paper • 2110.02178 • Published • 3
Mobile-VIT: Optimized for Qualcomm Devices
MobileVit 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 Mobile-VIT 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 Mobile-VIT 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 Mobile-VIT on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.image_classification
Model Stats:
- Model checkpoint: Imagenet
- Input resolution: 224x224
- Number of parameters: 5.57M
- Model size (float): 21.4 MB
- Model size (w8a16): 6.56 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| Mobile-VIT | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.665 ms | 1 - 102 MB | NPU |
| Mobile-VIT | ONNX | float | Snapdragon® X2 Elite | 1.894 ms | 12 - 12 MB | NPU |
| Mobile-VIT | ONNX | float | Snapdragon® X Elite | 3.952 ms | 12 - 12 MB | NPU |
| Mobile-VIT | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.526 ms | 0 - 109 MB | NPU |
| Mobile-VIT | ONNX | float | Qualcomm® QCS8550 (Proxy) | 3.591 ms | 0 - 17 MB | NPU |
| Mobile-VIT | ONNX | float | Qualcomm® QCS9075 | 4.693 ms | 1 - 4 MB | NPU |
| Mobile-VIT | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.976 ms | 0 - 88 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 1.919 ms | 0 - 114 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® X2 Elite | 2.068 ms | 7 - 7 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® X Elite | 5.06 ms | 8 - 8 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 3.118 ms | 0 - 131 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Qualcomm® QCS6490 | 313.39 ms | 63 - 67 MB | CPU |
| Mobile-VIT | ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 4.636 ms | 0 - 10 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Qualcomm® QCS9075 | 5.087 ms | 0 - 3 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Qualcomm® QCM6690 | 141.953 ms | 65 - 75 MB | CPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 2.336 ms | 0 - 83 MB | NPU |
| Mobile-VIT | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 122.831 ms | 65 - 76 MB | CPU |
| Mobile-VIT | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.608 ms | 1 - 72 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Snapdragon® X2 Elite | 2.136 ms | 1 - 1 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Snapdragon® X Elite | 3.849 ms | 1 - 1 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.462 ms | 0 - 99 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 9.92 ms | 1 - 68 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 3.491 ms | 1 - 2 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® SA8775P | 4.269 ms | 1 - 70 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® QCS9075 | 4.486 ms | 3 - 5 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 5.664 ms | 0 - 92 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® SA7255P | 9.92 ms | 1 - 68 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Qualcomm® SA8295P | 6.477 ms | 1 - 62 MB | NPU |
| Mobile-VIT | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.925 ms | 1 - 72 MB | NPU |
| Mobile-VIT | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.653 ms | 0 - 86 MB | NPU |
| Mobile-VIT | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.579 ms | 0 - 107 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 10.232 ms | 0 - 81 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 3.682 ms | 0 - 7 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® SA8775P | 4.47 ms | 0 - 83 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® QCS9075 | 4.581 ms | 0 - 15 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 6.039 ms | 0 - 103 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® SA7255P | 10.232 ms | 0 - 81 MB | NPU |
| Mobile-VIT | TFLITE | float | Qualcomm® SA8295P | 6.699 ms | 0 - 84 MB | NPU |
| Mobile-VIT | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 2.022 ms | 0 - 78 MB | NPU |
License
- The license for the original implementation of Mobile-VIT can be found here.
References
- MOBILEVIT: LIGHT-WEIGHT, GENERAL-PURPOSE, AND MOBILE-FRIENDLY VISION TRANSFORMER
- 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.