--- library_name: pytorch license: other tags: - backbone - bu_auto - android pipeline_tag: image-classification --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/web-assets/model_demo.png) # GPUNet: Optimized for Mobile Deployment ## Imagenet classifier and general purpose backbone GPUNet 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 model is an implementation of GPUNet found [here](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/GPUNet). This repository provides scripts to run GPUNet on Qualcomm® devices. More details on model performance across various devices, can be found [here](https://aihub.qualcomm.com/models/gpunet). ### Model Details - **Model Type:** Model_use_case.image_classification - **Model Stats:** - Model checkpoint: Imagenet - Input resolution: 224x224 - Number of parameters: 10.49M - Model size (float): 45.28MB - Model size (w8a8): 21.3MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | GPUNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 4.646 ms | 0 - 152 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 4.645 ms | 1 - 125 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 2.252 ms | 0 - 186 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 2.251 ms | 1 - 161 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.211 ms | 0 - 2 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.244 ms | 1 - 2 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.209 ms | 0 - 29 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.onnx.zip) | | GPUNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 7.046 ms | 0 - 152 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 1.704 ms | 1 - 126 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 4.646 ms | 0 - 152 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 4.645 ms | 1 - 125 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 2.213 ms | 0 - 159 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 2.228 ms | 1 - 132 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 7.046 ms | 0 - 152 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 1.704 ms | 1 - 126 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.897 ms | 0 - 183 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 0.9 ms | 1 - 159 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.881 ms | 0 - 128 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.onnx.zip) | | GPUNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.699 ms | 0 - 154 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.695 ms | 0 - 129 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.715 ms | 0 - 103 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.onnx.zip) | | GPUNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | TFLITE | 0.564 ms | 0 - 154 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.tflite) | | GPUNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | QNN_DLC | 0.569 ms | 1 - 129 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | ONNX | 0.619 ms | 0 - 100 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.onnx.zip) | | GPUNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 1.364 ms | 1 - 1 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.dlc) | | GPUNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1.122 ms | 24 - 24 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet.onnx.zip) | | GPUNet | w8a16 | Dragonwing Q-6690 MTP | Qualcomm® QCM6690 | QNN_DLC | 6.567 ms | 0 - 145 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Dragonwing Q-6690 MTP | Qualcomm® QCM6690 | ONNX | 53.54 ms | 26 - 40 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | Dragonwing RB3 Gen 2 Vision Kit | Qualcomm® QCS6490 | QNN_DLC | 3.208 ms | 0 - 2 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Dragonwing RB3 Gen 2 Vision Kit | Qualcomm® QCS6490 | ONNX | 101.582 ms | 19 - 34 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 2.515 ms | 0 - 134 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 1.445 ms | 0 - 158 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.067 ms | 0 - 2 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 1.009 ms | 0 - 16 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 5.508 ms | 0 - 134 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 2.515 ms | 0 - 134 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 1.648 ms | 0 - 140 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 5.508 ms | 0 - 134 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 0.77 ms | 0 - 158 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.703 ms | 0 - 137 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.527 ms | 0 - 139 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.537 ms | 0 - 113 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | QNN_DLC | 1.302 ms | 0 - 141 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | ONNX | 51.701 ms | 30 - 46 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | QNN_DLC | 0.442 ms | 0 - 138 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | ONNX | 0.482 ms | 0 - 115 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 1.242 ms | 0 - 0 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.dlc) | | GPUNet | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.973 ms | 12 - 12 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a16.onnx.zip) | | GPUNet | w8a8 | Dragonwing Q-6690 MTP | Qualcomm® QCM6690 | TFLITE | 3.026 ms | 0 - 137 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Dragonwing Q-6690 MTP | Qualcomm® QCM6690 | QNN_DLC | 3.55 ms | 0 - 138 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Dragonwing Q-6690 MTP | Qualcomm® QCM6690 | ONNX | 10.103 ms | 0 - 14 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | Dragonwing RB3 Gen 2 Vision Kit | Qualcomm® QCS6490 | TFLITE | 1.51 ms | 0 - 16 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Dragonwing RB3 Gen 2 Vision Kit | Qualcomm® QCS6490 | QNN_DLC | 2.003 ms | 2 - 4 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Dragonwing RB3 Gen 2 Vision Kit | Qualcomm® QCS6490 | ONNX | 16.955 ms | 4 - 18 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 1.12 ms | 0 - 131 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 1.435 ms | 0 - 132 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 0.685 ms | 0 - 155 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 0.875 ms | 0 - 157 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 0.432 ms | 0 - 3 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 0.623 ms | 0 - 2 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 0.797 ms | 0 - 2 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 0.626 ms | 0 - 132 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 0.808 ms | 0 - 132 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 1.12 ms | 0 - 131 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 1.435 ms | 0 - 132 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 0.85 ms | 0 - 137 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 1.052 ms | 0 - 138 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 0.626 ms | 0 - 132 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 0.808 ms | 0 - 132 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.34 ms | 0 - 156 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 0.465 ms | 0 - 158 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 0.57 ms | 0 - 132 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.275 ms | 0 - 136 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.34 ms | 0 - 136 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 0.501 ms | 0 - 111 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | TFLITE | 0.639 ms | 0 - 137 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | QNN_DLC | 0.803 ms | 0 - 138 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | ONNX | 9.278 ms | 10 - 25 MB | CPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | TFLITE | 0.242 ms | 0 - 137 MB | NPU | [GPUNet.tflite](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.tflite) | | GPUNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | QNN_DLC | 0.286 ms | 0 - 138 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen 5 Mobile | ONNX | 0.468 ms | 0 - 113 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | | GPUNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 0.717 ms | 0 - 0 MB | NPU | [GPUNet.dlc](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.dlc) | | GPUNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.714 ms | 12 - 12 MB | NPU | [GPUNet.onnx.zip](https://huggingface.co/qualcomm/GPUNet/blob/main/GPUNet_w8a8.onnx.zip) | ## Installation Install the package via pip: ```bash pip install qai-hub-models ``` ## Configure Qualcomm® AI Hub Workbench to run this model on a cloud-hosted device Sign-in to [Qualcomm® AI Hub Workbench](https://workbench.aihub.qualcomm.com/) with your Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`. With this API token, you can configure your client to run models on the cloud hosted devices. ```bash qai-hub configure --api_token API_TOKEN ``` Navigate to [docs](https://workbench.aihub.qualcomm.com/docs/) for more information. ## Demo off target The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input. ```bash python -m qai_hub_models.models.gpunet.demo ``` The above demo runs a reference implementation of pre-processing, model inference, and post processing. **NOTE**: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above). ``` %run -m qai_hub_models.models.gpunet.demo ``` ### Run model on a cloud-hosted device In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following: * Performance check on-device on a cloud-hosted device * Downloads compiled assets that can be deployed on-device for Android. * Accuracy check between PyTorch and on-device outputs. ```bash python -m qai_hub_models.models.gpunet.export ``` ## How does this work? This [export script](https://aihub.qualcomm.com/models/gpunet/qai_hub_models/models/GPUNet/export.py) leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model on-device. Lets go through each step below in detail: Step 1: **Compile model for on-device deployment** To compile a PyTorch model for on-device deployment, we first trace the model in memory using the `jit.trace` and then call the `submit_compile_job` API. ```python import torch import qai_hub as hub from qai_hub_models.models.gpunet import Model # Load the model torch_model = Model.from_pretrained() # Device device = hub.Device("Samsung Galaxy S25") # Trace model input_shape = torch_model.get_input_spec() sample_inputs = torch_model.sample_inputs() pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()]) # Compile model on a specific device compile_job = hub.submit_compile_job( model=pt_model, device=device, input_specs=torch_model.get_input_spec(), ) # Get target model to run on-device target_model = compile_job.get_target_model() ``` Step 2: **Performance profiling on cloud-hosted device** After compiling models from step 1. Models can be profiled model on-device using the `target_model`. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics. ```python profile_job = hub.submit_profile_job( model=target_model, device=device, ) ``` Step 3: **Verify on-device accuracy** To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device. ```python input_data = torch_model.sample_inputs() inference_job = hub.submit_inference_job( model=target_model, device=device, inputs=input_data, ) on_device_output = inference_job.download_output_data() ``` With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output. **Note**: This on-device profiling and inference requires access to Qualcomm® AI Hub Workbench. [Sign up for access](https://myaccount.qualcomm.com/signup). ## Deploying compiled model to Android The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): [This tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This [sample app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html) provides instructions on how to use the `.so` shared library in an Android application. ## View on Qualcomm® AI Hub Get more details on GPUNet's performance across various devices [here](https://aihub.qualcomm.com/models/gpunet). Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/) ## License * The license for the original implementation of GPUNet can be found [here](http://www.apache.org/licenses/LICENSE-2.0). ## References * [GPUNet: Searching the Deployable Convolution Neural Networks for GPUs](https://arxiv.org/abs/2205.00841) * [Source Model Implementation](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/GPUNet) ## 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).