README.md

---
library_name: pytorch
license: other
tags:
- backbone
- android
pipeline_tag: image-classification

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/web-assets/model_demo.png)

# RegNet: Optimized for Qualcomm Devices

RegNet 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 RegNet found [here](https://github.com/pytorch/vision/blob/main/torchvision/models/regnet.py).
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/main/qai_hub_models/models/regnet) 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 2.42, ONNX Runtime 1.24.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.43 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.43 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.43, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.43, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/regnet/releases/v0.50.0/regnet-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[RegNet on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/regnet)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/regnet) 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 [RegNet on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/qai_hub_models/models/regnet) for usage instructions.

## Model Details

**Model Type:** Model_use_case.image_classification

**Model Stats:**
- Model checkpoint: Imagenet
- Input resolution: 224x224
- Number of parameters: 15.3M
- Model size (float): 58.3 MB
- Model size (w8a8): 15.4 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| RegNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.839 ms | 1 - 88 MB | NPU
| RegNet | ONNX | float | Snapdragon® X2 Elite | 0.904 ms | 39 - 39 MB | NPU
| RegNet | ONNX | float | Snapdragon® X Elite | 1.957 ms | 39 - 39 MB | NPU
| RegNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 1.246 ms | 0 - 136 MB | NPU
| RegNet | ONNX | float | Qualcomm® QCS8550 (Proxy) | 1.749 ms | 0 - 44 MB | NPU
| RegNet | ONNX | float | Qualcomm® QCS9075 | 2.849 ms | 1 - 4 MB | NPU
| RegNet | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.0 ms | 0 - 85 MB | NPU
| RegNet | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.504 ms | 0 - 89 MB | NPU
| RegNet | ONNX | w8a8 | Snapdragon® X2 Elite | 0.502 ms | 20 - 20 MB | NPU
| RegNet | ONNX | w8a8 | Snapdragon® X Elite | 1.123 ms | 20 - 20 MB | NPU
| RegNet | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.654 ms | 0 - 135 MB | NPU
| RegNet | ONNX | w8a8 | Qualcomm® QCS6490 | 27.965 ms | 9 - 19 MB | CPU
| RegNet | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.901 ms | 0 - 28 MB | NPU
| RegNet | ONNX | w8a8 | Qualcomm® QCS9075 | 1.076 ms | 0 - 3 MB | NPU
| RegNet | ONNX | w8a8 | Qualcomm® QCM6690 | 17.962 ms | 8 - 17 MB | CPU
| RegNet | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.554 ms | 0 - 89 MB | NPU
| RegNet | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 13.724 ms | 8 - 18 MB | CPU
| RegNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.899 ms | 1 - 81 MB | NPU
| RegNet | QNN_DLC | float | Snapdragon® X2 Elite | 1.231 ms | 1 - 1 MB | NPU
| RegNet | QNN_DLC | float | Snapdragon® X Elite | 2.317 ms | 1 - 1 MB | NPU
| RegNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 1.426 ms | 0 - 128 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 9.898 ms | 1 - 76 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 2.073 ms | 1 - 161 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® SA8775P | 3.305 ms | 1 - 79 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® QCS9075 | 3.071 ms | 1 - 3 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 3.536 ms | 0 - 115 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® SA7255P | 9.898 ms | 1 - 76 MB | NPU
| RegNet | QNN_DLC | float | Qualcomm® SA8295P | 3.496 ms | 0 - 65 MB | NPU
| RegNet | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.118 ms | 0 - 77 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.44 ms | 0 - 80 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.625 ms | 0 - 0 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 1.077 ms | 0 - 0 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.63 ms | 0 - 108 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 2.699 ms | 0 - 2 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 2.263 ms | 0 - 76 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.893 ms | 0 - 2 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 1.32 ms | 0 - 79 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 1.084 ms | 2 - 4 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 7.108 ms | 0 - 195 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 1.293 ms | 0 - 110 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 2.263 ms | 0 - 76 MB | NPU
| RegNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 1.608 ms | 0 - 75 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.497 ms | 0 - 79 MB | NPU
| RegNet | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.171 ms | 0 - 77 MB | NPU
| RegNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.9 ms | 0 - 105 MB | NPU
| RegNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 1.396 ms | 0 - 156 MB | NPU
| RegNet | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 9.871 ms | 0 - 100 MB | NPU
| RegNet | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 2.092 ms | 0 - 3 MB | NPU
| RegNet | TFLITE | float | Qualcomm® SA8775P | 3.376 ms | 0 - 103 MB | NPU
| RegNet | TFLITE | float | Qualcomm® QCS9075 | 3.054 ms | 0 - 42 MB | NPU
| RegNet | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 3.455 ms | 0 - 143 MB | NPU
| RegNet | TFLITE | float | Qualcomm® SA7255P | 9.871 ms | 0 - 100 MB | NPU
| RegNet | TFLITE | float | Qualcomm® SA8295P | 3.5 ms | 0 - 83 MB | NPU
| RegNet | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.109 ms | 0 - 102 MB | NPU
| RegNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.384 ms | 0 - 78 MB | NPU
| RegNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.516 ms | 0 - 111 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCS6490 | 2.326 ms | 0 - 21 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 1.984 ms | 0 - 74 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.726 ms | 0 - 3 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® SA8775P | 1.139 ms | 0 - 77 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCS9075 | 0.892 ms | 0 - 22 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCM6690 | 6.661 ms | 0 - 190 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 1.131 ms | 0 - 116 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® SA7255P | 1.984 ms | 0 - 74 MB | NPU
| RegNet | TFLITE | w8a8 | Qualcomm® SA8295P | 1.414 ms | 0 - 71 MB | NPU
| RegNet | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.436 ms | 0 - 70 MB | NPU
| RegNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.991 ms | 0 - 71 MB | NPU

## License
* The license for the original implementation of RegNet can be found
  [here](https://github.com/pytorch/vision/blob/main/LICENSE).

## References
* [Designing Network Design Spaces](https://arxiv.org/abs/2003.13678)
* [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/regnet.py)

## 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).