SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size
Paper • 1602.07360 • Published • 1
SqueezeNet-1.1: Optimized for Qualcomm Devices
SqueezeNet 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 SqueezeNet-1.1 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 1.25.0 | Download |
| ONNX | w8a8 | Universal | QAIRT 2.45, ONNX Runtime 1.25.0 | Download |
| QNN_DLC | float | Universal | QAIRT 2.45 | Download |
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | Download |
| TFLITE | float | Universal | QAIRT 2.45 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit SqueezeNet-1.1 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 SqueezeNet-1.1 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: 1.24M
- Model size (float): 4.73 MB
- Model size (w8a8): 1.30 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| SqueezeNet-1.1 | ONNX | float | Snapdragon® X2 Elite | 0.183 ms | 212 - 212 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Snapdragon® X Elite | 0.379 ms | 148 - 148 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.271 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 0.572 ms | 1 - 29 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Qualcomm® QCS8550 (Proxy) | 0.381 ms | 0 - 72 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Qualcomm® QCS8450 | 0.572 ms | 1 - 29 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Snapdragon® 8 Elite Mobile | 0.223 ms | 0 - 25 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.188 ms | 0 - 17 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Qualcomm® QCS9075 | 0.649 ms | 0 - 47 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Qualcomm® QCS8750 | 0.223 ms | 0 - 25 MB | NPU |
| SqueezeNet-1.1 | ONNX | float | Qualcomm® QCS7181 | 0.379 ms | 148 - 148 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® X2 Elite | 0.217 ms | 181 - 181 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® X Elite | 0.37 ms | 149 - 149 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.271 ms | 0 - 22 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.45 ms | 0 - 30 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS6490 | 0.817 ms | 0 - 47 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.379 ms | 0 - 4 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS8450 | 0.45 ms | 0 - 30 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.217 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.349 ms | 0 - 18 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCM6690 | 1.183 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS9075 | 0.507 ms | 0 - 47 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Snapdragon® 8 Elite Mobile | 0.235 ms | 0 - 23 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS7790 | 0.349 ms | 0 - 18 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS8750 | 0.235 ms | 0 - 23 MB | NPU |
| SqueezeNet-1.1 | ONNX | w8a8 | Qualcomm® QCS7181 | 0.37 ms | 149 - 149 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® X2 Elite | 0.348 ms | 1 - 1 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® X Elite | 0.788 ms | 1 - 1 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.446 ms | 0 - 31 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 1.259 ms | 0 - 37 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS8275 | 2.096 ms | 1 - 22 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 0.654 ms | 1 - 3 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS8450 | 1.259 ms | 0 - 37 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 0.328 ms | 0 - 26 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® SA8295P | 1.126 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.242 ms | 1 - 23 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® SA7255P | 2.096 ms | 1 - 22 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS9075 | 0.867 ms | 1 - 3 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS8750 | 0.328 ms | 0 - 26 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | float | Qualcomm® QCS7181 | 0.788 ms | 1 - 1 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.234 ms | 0 - 0 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® X Elite | 0.484 ms | 0 - 0 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.269 ms | 0 - 30 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.43 ms | 0 - 35 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 1.054 ms | 0 - 2 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS8275 | 0.918 ms | 0 - 22 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.369 ms | 0 - 28 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 0.43 ms | 0 - 35 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.153 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.369 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 1.445 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 0.442 ms | 0 - 2 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® SA7255P | 0.918 ms | 0 - 22 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 0.181 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® SA8295P | 0.701 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS7790 | 0.369 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS8750 | 0.181 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | QNN_DLC | w8a8 | Qualcomm® QCS7181 | 0.484 ms | 0 - 0 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.446 ms | 0 - 33 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 1.265 ms | 0 - 39 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® QCS8275 | 2.098 ms | 0 - 24 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 0.65 ms | 0 - 57 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® SA8775P | 3.022 ms | 0 - 27 MB | GPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® SA8650P | 3.022 ms | 0 - 27 MB | GPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® SA8255P | 3.022 ms | 0 - 27 MB | GPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® QCS8450 | 1.265 ms | 0 - 39 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Snapdragon® 8 Elite Mobile | 0.339 ms | 0 - 24 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® SA8295P | 1.154 ms | 0 - 21 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.236 ms | 0 - 25 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® SA7255P | 2.098 ms | 0 - 24 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® QCS9075 | 0.874 ms | 0 - 5 MB | NPU |
| SqueezeNet-1.1 | TFLITE | float | Qualcomm® QCS8750 | 0.339 ms | 0 - 24 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.147 ms | 0 - 30 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.244 ms | 0 - 35 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS6490 | 0.608 ms | 0 - 3 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS8275 | 0.615 ms | 0 - 21 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.198 ms | 0 - 3 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® SA8775P | 3.151 ms | 0 - 26 MB | GPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® SA8650P | 3.151 ms | 0 - 26 MB | GPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® SA8255P | 3.151 ms | 0 - 26 MB | GPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS8450 | 0.244 ms | 0 - 35 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.095 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.201 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCM6690 | 0.936 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS9075 | 0.275 ms | 0 - 3 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® SA7255P | 0.615 ms | 0 - 21 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 0.113 ms | 0 - 20 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® SA8295P | 0.48 ms | 0 - 17 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS7790 | 0.201 ms | 0 - 19 MB | NPU |
| SqueezeNet-1.1 | TFLITE | w8a8 | Qualcomm® QCS8750 | 0.113 ms | 0 - 20 MB | NPU |
License
- The license for the original implementation of SqueezeNet-1.1 can be found here.
References
- SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size
- 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.