--- library_name: pytorch license: apache-2.0 tags: - backbone - android pipeline_tag: image-classification --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/repvit/web-assets/model_demo.png) # RepViT: Optimized for Qualcomm Devices RepViT is a lightweight pure CNN model for mobile devices, incorporating efficient architectural designs from Vision Transformers into CNNs. It achieves over 80% top-1 accuracy on ImageNet with 1ms latency on an iPhone 12. This is based on the implementation of RepViT found [here](https://github.com/THU-MIG/RepViT). This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/repvit) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary). Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) 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 2.50, ONNX Runtime 1.30.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/repvit/releases/v0.64.0/repvit-onnx-float.zip) | QNN_DLC | float | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/repvit/releases/v0.64.0/repvit-qnn_dlc-float.zip) | TFLITE | float | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/repvit/releases/v0.64.0/repvit-tflite-float.zip) For more device-specific assets and performance metrics, visit **[RepViT on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/repvit)**. ### Option 2: Export with Custom Configurations Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/repvit) 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 [RepViT on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.64.0/src/qai_hub_models/models/repvit) for usage instructions. ## Model Details **Model Type:** Model_use_case.image_classification **Model Stats:** - Input resolution: 224x224 - Model checkpoint: Imagenet - Model size (float): 91 MB - Number of parameters: 22.9M ## Performance Summary | Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |---|---|---|---|---|---|--- | RepViT | ONNX | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.427 ms | 0 - 114 MB | NPU | RepViT | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.775 ms | 0 - 113 MB | NPU | RepViT | ONNX | float | Snapdragon® X2 Elite | 1.68 ms | 2 - 2 MB | NPU | RepViT | ONNX | float | Snapdragon® X Elite | 3.34 ms | 48 - 48 MB | NPU | RepViT | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.357 ms | 0 - 195 MB | NPU | RepViT | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 11.094 ms | 0 - 190 MB | NPU | RepViT | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 4.604 ms | 0 - 5 MB | NPU | RepViT | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.173 ms | 0 - 50 MB | NPU | RepViT | ONNX | float | Qualcomm® QCS8450 | 11.094 ms | 0 - 190 MB | NPU | RepViT | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 4.341 ms | 0 - 4 MB | NPU | RepViT | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 3.34 ms | 48 - 48 MB | NPU | RepViT | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 1.775 ms | 0 - 113 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.512 ms | 1 - 105 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.886 ms | 1 - 102 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® X2 Elite | 2.022 ms | 1 - 1 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® X Elite | 3.834 ms | 1 - 1 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.517 ms | 0 - 180 MB | NPU | RepViT | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 11.873 ms | 1 - 174 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 4.738 ms | 1 - 4 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.477 ms | 1 - 154 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® SA8775P | 4.719 ms | 1 - 103 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® SA8650P | 4.719 ms | 1 - 103 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® SA8255P | 4.719 ms | 1 - 103 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® QCS8450 | 11.873 ms | 1 - 174 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 4.557 ms | 3 - 5 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 3.834 ms | 1 - 1 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 1.886 ms | 1 - 102 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® SA7255P | 14.337 ms | 1 - 101 MB | NPU | RepViT | QNN_DLC | float | Qualcomm® SA8295P | 9.547 ms | 1 - 103 MB | NPU | RepViT | TFLITE | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.519 ms | 0 - 171 MB | NPU | RepViT | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.873 ms | 0 - 160 MB | NPU | RepViT | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.535 ms | 0 - 231 MB | NPU | RepViT | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 11.918 ms | 0 - 223 MB | NPU | RepViT | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 4.751 ms | 0 - 51 MB | NPU | RepViT | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.425 ms | 0 - 9 MB | NPU | RepViT | TFLITE | float | Qualcomm® SA8775P | 4.739 ms | 0 - 163 MB | NPU | RepViT | TFLITE | float | Qualcomm® SA8650P | 4.739 ms | 0 - 163 MB | NPU | RepViT | TFLITE | float | Qualcomm® SA8255P | 4.739 ms | 0 - 163 MB | NPU | RepViT | TFLITE | float | Qualcomm® QCS8450 | 11.918 ms | 0 - 223 MB | NPU | RepViT | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 4.561 ms | 0 - 51 MB | NPU | RepViT | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 1.873 ms | 0 - 160 MB | NPU | RepViT | TFLITE | float | Qualcomm® SA8295P | 9.493 ms | 0 - 160 MB | NPU ## License * The license for the original implementation of RepViT can be found [here](https://github.com/THU-MIG/RepViT/blob/main/LICENSE). ## References * [RepViT: Revisiting Mobile CNN From ViT Perspective](https://arxiv.org/abs/2307.09283) * [Source Model Implementation](https://github.com/THU-MIG/RepViT) ## Community * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI. * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).