|
|
--- |
|
|
library_name: pytorch |
|
|
license: other |
|
|
tags: |
|
|
- backbone |
|
|
- bu_auto |
|
|
- android |
|
|
pipeline_tag: image-classification |
|
|
|
|
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
# EfficientNet-B4: Optimized for Qualcomm Devices |
|
|
|
|
|
EfficientNetB4 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-B4 found [here](https://github.com/pytorch/vision/blob/main/torchvision/models/efficientnet.py). |
|
|
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/efficientnet_b4) 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 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/efficientnet_b4/releases/v0.46.0/efficientnet_b4-onnx-float.zip) |
|
|
| ONNX | w8a16 | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/efficientnet_b4/releases/v0.46.0/efficientnet_b4-onnx-w8a16.zip) |
|
|
| QNN_DLC | float | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/efficientnet_b4/releases/v0.46.0/efficientnet_b4-qnn_dlc-float.zip) |
|
|
| QNN_DLC | w8a16 | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/efficientnet_b4/releases/v0.46.0/efficientnet_b4-qnn_dlc-w8a16.zip) |
|
|
| TFLITE | float | Universal | QAIRT 2.42, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/efficientnet_b4/releases/v0.46.0/efficientnet_b4-tflite-float.zip) |
|
|
|
|
|
For more device-specific assets and performance metrics, visit **[EfficientNet-B4 on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/efficientnet_b4)**. |
|
|
|
|
|
|
|
|
### Option 2: Export with Custom Configurations |
|
|
|
|
|
Use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/efficientnet_b4) 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-B4 on GitHub](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/efficientnet_b4) for usage instructions. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
**Model Type:** Model_use_case.image_classification |
|
|
|
|
|
**Model Stats:** |
|
|
- Model checkpoint: Imagenet |
|
|
- Input resolution: 380x380 |
|
|
- Number of parameters: 19.3M |
|
|
- Model size (float): 73.6 MB |
|
|
- Model size (w8a16): 24.0 MB |
|
|
|
|
|
## Performance Summary |
|
|
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|
|
|---|---|---|---|---|---|--- |
|
|
| EfficientNet-B4 | ONNX | float | Snapdragon® X Elite | 3.25 ms | 45 - 45 MB | NPU |
|
|
| EfficientNet-B4 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 2.451 ms | 0 - 186 MB | NPU |
|
|
| EfficientNet-B4 | ONNX | float | Qualcomm® QCS8550 (Proxy) | 3.185 ms | 0 - 107 MB | NPU |
|
|
| EfficientNet-B4 | ONNX | float | Qualcomm® QCS9075 | 4.16 ms | 0 - 4 MB | NPU |
|
|
| EfficientNet-B4 | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.933 ms | 0 - 135 MB | NPU |
|
|
| EfficientNet-B4 | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.656 ms | 0 - 137 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Snapdragon® X Elite | 3.598 ms | 1 - 1 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.417 ms | 0 - 124 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 12.05 ms | 1 - 68 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 3.315 ms | 1 - 2 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Qualcomm® QCS9075 | 4.193 ms | 1 - 3 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 7.805 ms | 0 - 145 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.858 ms | 1 - 74 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.507 ms | 0 - 73 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Snapdragon® X Elite | 3.786 ms | 0 - 0 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 2.328 ms | 0 - 151 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCS6490 | 8.737 ms | 2 - 4 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCS8275 (Proxy) | 6.559 ms | 0 - 99 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCS8550 (Proxy) | 3.429 ms | 0 - 2 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCS9075 | 3.79 ms | 0 - 2 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCM6690 | 17.274 ms | 0 - 229 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Qualcomm® QCS8450 (Proxy) | 4.203 ms | 0 - 153 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 1.599 ms | 0 - 102 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 3.621 ms | 0 - 107 MB | NPU |
|
|
| EfficientNet-B4 | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 1.319 ms | 0 - 102 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.421 ms | 0 - 168 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 12.112 ms | 0 - 106 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 3.326 ms | 0 - 3 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Qualcomm® QCS9075 | 4.179 ms | 0 - 48 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 7.901 ms | 0 - 186 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.857 ms | 0 - 111 MB | NPU |
|
|
| EfficientNet-B4 | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.508 ms | 0 - 110 MB | NPU |
|
|
|
|
|
## License |
|
|
* The license for the original implementation of EfficientNet-B4 can be found |
|
|
[here](https://github.com/pytorch/vision/blob/main/LICENSE). |
|
|
|
|
|
## References |
|
|
* [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) |
|
|
* [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/efficientnet.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). |
|
|
|
