--- library_name: pytorch license: other tags: - android pipeline_tag: keypoint-detection --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/hrnet_pose/web-assets/model_demo.png) # HRNetPose: Optimized for Mobile Deployment ## Perform accurate human pose estimation HRNet performs pose estimation in high-resolution representations. This model is an implementation of HRNetPose found [here](https://github.com/leoxiaobin/deep-high-resolution-net.pytorch). This repository provides scripts to run HRNetPose on Qualcomm® devices. More details on model performance across various devices, can be found [here](https://aihub.qualcomm.com/models/hrnet_pose). ### Model Details - **Model Type:** Model_use_case.pose_estimation - **Model Stats:** - Model checkpoint: hrnet_posenet_FP32_state_dict - Input resolution: 256x192 - Number of parameters: 28.5M - Model size (float): 109 MB - Model size (w8a8): 28 MB | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | HRNetPose | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 14.4 ms | 0 - 71 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN | 14.249 ms | 1 - 10 MB | NPU | Use Export Script | | HRNetPose | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 3.877 ms | 0 - 118 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN | 5.124 ms | 0 - 52 MB | NPU | Use Export Script | | HRNetPose | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 2.659 ms | 0 - 77 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN | 2.649 ms | 1 - 3 MB | NPU | Use Export Script | | HRNetPose | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 4.461 ms | 0 - 71 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN | 4.344 ms | 1 - 12 MB | NPU | Use Export Script | | HRNetPose | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 14.4 ms | 0 - 71 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | SA7255P ADP | Qualcomm® SA7255P | QNN | 14.249 ms | 1 - 10 MB | NPU | Use Export Script | | HRNetPose | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 2.684 ms | 0 - 79 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN | 2.677 ms | 1 - 2 MB | NPU | Use Export Script | | HRNetPose | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 4.57 ms | 0 - 67 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | SA8295P ADP | Qualcomm® SA8295P | QNN | 4.537 ms | 1 - 18 MB | NPU | Use Export Script | | HRNetPose | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 2.68 ms | 0 - 19 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN | 2.663 ms | 1 - 3 MB | NPU | Use Export Script | | HRNetPose | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 4.461 ms | 0 - 71 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | SA8775P ADP | Qualcomm® SA8775P | QNN | 4.344 ms | 1 - 12 MB | NPU | Use Export Script | | HRNetPose | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 2.678 ms | 0 - 78 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN | 2.646 ms | 0 - 16 MB | NPU | Use Export Script | | HRNetPose | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 2.715 ms | 0 - 162 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.onnx) | | HRNetPose | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 1.965 ms | 0 - 116 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN | 1.997 ms | 1 - 53 MB | NPU | Use Export Script | | HRNetPose | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 2.084 ms | 0 - 89 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.onnx) | | HRNetPose | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 1.853 ms | 0 - 74 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.tflite) | | HRNetPose | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN | 1.808 ms | 0 - 40 MB | NPU | Use Export Script | | HRNetPose | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 1.943 ms | 0 - 52 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.onnx) | | HRNetPose | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 2.823 ms | 1 - 1 MB | NPU | Use Export Script | | HRNetPose | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.699 ms | 55 - 55 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose.onnx) | | HRNetPose | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 2.665 ms | 0 - 48 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN | 2.888 ms | 0 - 9 MB | NPU | Use Export Script | | HRNetPose | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.313 ms | 0 - 91 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN | 1.92 ms | 0 - 78 MB | NPU | Use Export Script | | HRNetPose | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 0.996 ms | 0 - 133 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN | 1.146 ms | 0 - 2 MB | NPU | Use Export Script | | HRNetPose | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 1.326 ms | 0 - 50 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN | 1.422 ms | 0 - 15 MB | NPU | Use Export Script | | HRNetPose | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 3.813 ms | 0 - 75 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | QNN | 5.224 ms | 0 - 14 MB | NPU | Use Export Script | | HRNetPose | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 17.222 ms | 0 - 2 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 2.665 ms | 0 - 48 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN | 2.888 ms | 0 - 9 MB | NPU | Use Export Script | | HRNetPose | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 0.979 ms | 0 - 135 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN | 1.152 ms | 0 - 3 MB | NPU | Use Export Script | | HRNetPose | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 1.719 ms | 0 - 51 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN | 1.865 ms | 0 - 17 MB | NPU | Use Export Script | | HRNetPose | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 0.992 ms | 0 - 132 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN | 1.137 ms | 0 - 3 MB | NPU | Use Export Script | | HRNetPose | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 1.326 ms | 0 - 50 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN | 1.422 ms | 0 - 15 MB | NPU | Use Export Script | | HRNetPose | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 0.987 ms | 0 - 135 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN | 1.16 ms | 0 - 41 MB | NPU | Use Export Script | | HRNetPose | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 4.837 ms | 3 - 82 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.onnx) | | HRNetPose | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.743 ms | 0 - 89 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN | 0.822 ms | 0 - 80 MB | NPU | Use Export Script | | HRNetPose | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.744 ms | 0 - 162 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.onnx) | | HRNetPose | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 0.701 ms | 0 - 52 MB | NPU | [HRNetPose.tflite](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.tflite) | | HRNetPose | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN | 0.735 ms | 0 - 52 MB | NPU | Use Export Script | | HRNetPose | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 3.769 ms | 0 - 121 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.onnx) | | HRNetPose | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 1.26 ms | 0 - 0 MB | NPU | Use Export Script | | HRNetPose | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 5.901 ms | 27 - 27 MB | NPU | [HRNetPose.onnx](https://huggingface.co/qualcomm/HRNetPose/blob/main/HRNetPose_w8a8.onnx) | ## Installation Install the package via pip: ```bash pip install "qai-hub-models[hrnet-pose]" torch==2.4.1 -f https://download.openmmlab.com/mmcv/dist/cpu/torch2.4/index.html -f https://qaihub-public-python-wheels.s3.us-west-2.amazonaws.com/index.html ``` ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device Sign-in to [Qualcomm® AI Hub](https://app.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://app.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.hrnet_pose.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.hrnet_pose.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.hrnet_pose.export ``` ``` Profiling Results ------------------------------------------------------------ HRNetPose Device : cs_8275 (ANDROID 14) Runtime : TFLITE Estimated inference time (ms) : 14.4 Estimated peak memory usage (MB): [0, 71] Total # Ops : 516 Compute Unit(s) : npu (516 ops) gpu (0 ops) cpu (0 ops) ``` ## How does this work? This [export script](https://aihub.qualcomm.com/models/hrnet_pose/qai_hub_models/models/HRNetPose/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.hrnet_pose import Model # Load the model torch_model = Model.from_pretrained() # Device device = hub.Device("Samsung Galaxy S24") # 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. [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.hrnet_pose.demo --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.hrnet_pose.demo -- --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 HRNetPose's performance across various devices [here](https://aihub.qualcomm.com/models/hrnet_pose). Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/) ## License * The license for the original implementation of HRNetPose can be found [here](https://github.com/leoxiaobin/deep-high-resolution-net.pytorch/blob/master/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 * [Deep High-Resolution Representation Learning for Human Pose Estimation](https://arxiv.org/abs/1902.09212) * [Source Model Implementation](https://github.com/leoxiaobin/deep-high-resolution-net.pytorch) ## 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).