--- 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/swin_base/web-assets/model_demo.png) # Swin-Base: Optimized for Mobile Deployment ## Imagenet classifier and general purpose backbone SwinBase 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 Swin-Base found [here](https://github.com/pytorch/vision/blob/main/torchvision/models/swin_transformer.py). This repository provides scripts to run Swin-Base on Qualcomm® devices. More details on model performance across various devices, can be found [here](https://aihub.qualcomm.com/models/swin_base). ### Model Details - **Model Type:** Model_use_case.image_classification - **Model Stats:** - Model checkpoint: Imagenet - Input resolution: 224x224 - Number of parameters: 88.8M - Model size (float): 339 MB - Model size (w8a16): 90.2 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | Swin-Base | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 60.881 ms | 0 - 357 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 54.521 ms | 1 - 314 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 28.192 ms | 0 - 355 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 28.389 ms | 0 - 334 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 21.723 ms | 0 - 32 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 19.277 ms | 0 - 68 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 19.159 ms | 1 - 56 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.onnx.zip) | | Swin-Base | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 25.116 ms | 0 - 357 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 22.069 ms | 0 - 316 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 60.881 ms | 0 - 357 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 54.521 ms | 1 - 314 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 21.618 ms | 0 - 45 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 19.299 ms | 0 - 69 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 31.874 ms | 0 - 348 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 28.414 ms | 0 - 309 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 21.717 ms | 0 - 38 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 19.364 ms | 0 - 52 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 25.116 ms | 0 - 357 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 22.069 ms | 0 - 316 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 15.006 ms | 46 - 408 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 13.032 ms | 1 - 346 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 12.846 ms | 1 - 345 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.onnx.zip) | | Swin-Base | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 11.845 ms | 0 - 351 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 9.895 ms | 1 - 313 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 9.591 ms | 1 - 314 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.onnx.zip) | | Swin-Base | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 9.823 ms | 0 - 351 MB | NPU | [Swin-Base.tflite](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.tflite) | | Swin-Base | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 8.018 ms | 0 - 330 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 8.042 ms | 0 - 327 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.onnx.zip) | | Swin-Base | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 19.906 ms | 1035 - 1035 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.dlc) | | Swin-Base | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 19.472 ms | 175 - 175 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base.onnx.zip) | | Swin-Base | w8a16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 38.0 ms | 0 - 264 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 20.045 ms | 0 - 67 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 20.221 ms | 0 - 264 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 553.678 ms | 142 - 170 MB | CPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.onnx.zip) | | Swin-Base | w8a16 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 562.489 ms | 135 - 156 MB | CPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.onnx.zip) | | Swin-Base | w8a16 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 38.0 ms | 0 - 264 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 19.967 ms | 0 - 67 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 19.984 ms | 0 - 74 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 20.221 ms | 0 - 264 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 13.369 ms | 0 - 270 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 10.371 ms | 0 - 259 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 207.429 ms | 93 - 166 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.onnx.zip) | | Swin-Base | w8a16 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | QNN_DLC | 23.508 ms | 0 - 321 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Snapdragon 7 Gen 4 QRD | Snapdragon® 7 Gen 4 Mobile | ONNX | 595.51 ms | 119 - 141 MB | CPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.onnx.zip) | | Swin-Base | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 8.022 ms | 4 - 277 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 21.391 ms | 415 - 415 MB | NPU | [Swin-Base.dlc](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.dlc) | | Swin-Base | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 197.49 ms | 133 - 133 MB | NPU | [Swin-Base.onnx.zip](https://huggingface.co/qualcomm/Swin-Base/blob/main/Swin-Base_w8a16.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.swin_base.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.swin_base.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.swin_base.export ``` ## How does this work? This [export script](https://aihub.qualcomm.com/models/swin_base/qai_hub_models/models/Swin-Base/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.swin_base 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). ## Run demo on a cloud-hosted device You can also run the demo on-device. ```bash python -m qai_hub_models.models.swin_base.demo --eval-mode on-device ``` **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.swin_base.demo -- --eval-mode on-device ``` ## 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 Swin-Base's performance across various devices [here](https://aihub.qualcomm.com/models/swin_base). Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/) ## License * The license for the original implementation of Swin-Base can be found [here](https://github.com/pytorch/vision/blob/main/LICENSE). * The license for the compiled assets for on-device deployment can be found [here](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/Qualcomm+AI+Hub+Proprietary+License.pdf) ## References * [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) * [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/swin_transformer.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).