SINet: Optimized for Mobile Deployment
Lightweight portrait segmentation for background removal
SINet is a machine learning model that is designed to segment people from close-up portrait images in real time.
This model is an implementation of SINet found here.
This repository provides scripts to run SINet on Qualcomm® devices. More details on model performance across various devices, can be found here.
Model Details
- Model Type: Model_use_case.semantic_segmentation
- Model Stats:
- Model checkpoint: SINet.pth
- Input resolution: 224x224
- Number of output classes: 2 (foreground / background)
- Number of parameters: 91.9K
- Model size (float): 415 KB
- Model size (w8a8): 241 KB
| Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |
|---|---|---|---|---|---|---|---|---|
| SINet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 3.641 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 3.462 ms | 0 - 21 MB | NPU | SINet.dlc |
| SINet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.794 ms | 0 - 32 MB | NPU | SINet.tflite |
| SINet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 2.33 ms | 1 - 33 MB | NPU | SINet.dlc |
| SINet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.632 ms | 0 - 7 MB | NPU | SINet.tflite |
| SINet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.597 ms | 1 - 9 MB | NPU | SINet.dlc |
| SINet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 2.112 ms | 0 - 9 MB | NPU | SINet.onnx.zip |
| SINet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 2.107 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 2.025 ms | 1 - 22 MB | NPU | SINet.dlc |
| SINet | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 3.641 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 3.462 ms | 0 - 21 MB | NPU | SINet.dlc |
| SINet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 1.628 ms | 0 - 6 MB | NPU | SINet.tflite |
| SINet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 1.6 ms | 1 - 7 MB | NPU | SINet.dlc |
| SINet | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 2.378 ms | 0 - 26 MB | NPU | SINet.tflite |
| SINet | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 2.384 ms | 1 - 30 MB | NPU | SINet.dlc |
| SINet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 1.642 ms | 0 - 7 MB | NPU | SINet.tflite |
| SINet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 1.596 ms | 1 - 8 MB | NPU | SINet.dlc |
| SINet | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 2.107 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 2.025 ms | 1 - 22 MB | NPU | SINet.dlc |
| SINet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 1.085 ms | 0 - 35 MB | NPU | SINet.tflite |
| SINet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 1.068 ms | 1 - 35 MB | NPU | SINet.dlc |
| SINet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 1.363 ms | 0 - 34 MB | NPU | SINet.onnx.zip |
| SINet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.813 ms | 0 - 23 MB | NPU | SINet.tflite |
| SINet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.775 ms | 1 - 25 MB | NPU | SINet.dlc |
| SINet | float | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 1.102 ms | 0 - 25 MB | NPU | SINet.onnx.zip |
| SINet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.681 ms | 0 - 22 MB | NPU | SINet.tflite |
| SINet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 0.646 ms | 1 - 27 MB | NPU | SINet.dlc |
| SINet | float | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 1.029 ms | 1 - 30 MB | NPU | SINet.onnx.zip |
| SINet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 1.833 ms | 0 - 0 MB | NPU | SINet.dlc |
| SINet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.158 ms | 2 - 2 MB | NPU | SINet.onnx.zip |
| SINet | w8a16 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 3.495 ms | 0 - 23 MB | NPU | SINet.dlc |
| SINet | w8a16 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 2.37 ms | 0 - 33 MB | NPU | SINet.dlc |
| SINet | w8a16 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.808 ms | 0 - 8 MB | NPU | SINet.dlc |
| SINet | w8a16 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 2.151 ms | 0 - 23 MB | NPU | SINet.dlc |
| SINet | w8a16 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 3.495 ms | 0 - 23 MB | NPU | SINet.dlc |
| SINet | w8a16 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 1.803 ms | 0 - 9 MB | NPU | SINet.dlc |
| SINet | w8a16 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 2.875 ms | 0 - 30 MB | NPU | SINet.dlc |
| SINet | w8a16 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 1.818 ms | 0 - 8 MB | NPU | SINet.dlc |
| SINet | w8a16 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 2.151 ms | 0 - 23 MB | NPU | SINet.dlc |
| SINet | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 1.225 ms | 0 - 39 MB | NPU | SINet.dlc |
| SINet | w8a16 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.901 ms | 0 - 31 MB | NPU | SINet.dlc |
| SINet | w8a16 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 0.733 ms | 0 - 26 MB | NPU | SINet.dlc |
| SINet | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 2.019 ms | 6 - 6 MB | NPU | SINet.dlc |
| SINet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 2.271 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 2.565 ms | 0 - 22 MB | NPU | SINet.dlc |
| SINet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.231 ms | 0 - 37 MB | NPU | SINet.tflite |
| SINet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 1.449 ms | 0 - 33 MB | NPU | SINet.dlc |
| SINet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.127 ms | 0 - 8 MB | NPU | SINet.tflite |
| SINet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.297 ms | 0 - 7 MB | NPU | SINet.dlc |
| SINet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | ONNX | 5.259 ms | 4 - 14 MB | NPU | SINet.onnx.zip |
| SINet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 1.419 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 1.569 ms | 0 - 22 MB | NPU | SINet.dlc |
| SINet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 20.026 ms | 0 - 25 MB | NPU | SINet.tflite |
| SINet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | ONNX | 12.248 ms | 6 - 19 MB | CPU | SINet.onnx.zip |
| SINet | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | ONNX | 10.734 ms | 6 - 10 MB | CPU | SINet.onnx.zip |
| SINet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 2.271 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 2.565 ms | 0 - 22 MB | NPU | SINet.dlc |
| SINet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 1.123 ms | 0 - 8 MB | NPU | SINet.tflite |
| SINet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 1.295 ms | 0 - 8 MB | NPU | SINet.dlc |
| SINet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 1.695 ms | 0 - 27 MB | NPU | SINet.tflite |
| SINet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 1.919 ms | 0 - 28 MB | NPU | SINet.dlc |
| SINet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 1.122 ms | 0 - 8 MB | NPU | SINet.tflite |
| SINet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 1.281 ms | 0 - 8 MB | NPU | SINet.dlc |
| SINet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 1.419 ms | 0 - 21 MB | NPU | SINet.tflite |
| SINet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 1.569 ms | 0 - 22 MB | NPU | SINet.dlc |
| SINet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.78 ms | 0 - 33 MB | NPU | SINet.tflite |
| SINet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 0.895 ms | 0 - 34 MB | NPU | SINet.dlc |
| SINet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.603 ms | 0 - 25 MB | NPU | SINet.onnx.zip |
| SINet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | TFLITE | 0.592 ms | 0 - 25 MB | NPU | SINet.tflite |
| SINet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | QNN_DLC | 0.651 ms | 0 - 26 MB | NPU | SINet.dlc |
| SINet | w8a8 | Samsung Galaxy S25 | Snapdragon® 8 Elite For Galaxy Mobile | ONNX | 2.963 ms | 0 - 22 MB | NPU | SINet.onnx.zip |
| SINet | w8a8 | Snapdragon 7 Gen 5 QRD | Snapdragon® 7 Gen 5 Mobile | TFLITE | 5.432 ms | 0 - 24 MB | NPU | SINet.tflite |
| SINet | w8a8 | Snapdragon 7 Gen 5 QRD | Snapdragon® 7 Gen 5 Mobile | ONNX | 11.186 ms | 7 - 24 MB | CPU | SINet.onnx.zip |
| SINet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | TFLITE | 0.496 ms | 0 - 23 MB | NPU | SINet.tflite |
| SINet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | QNN_DLC | 0.528 ms | 0 - 27 MB | NPU | SINet.dlc |
| SINet | w8a8 | Snapdragon 8 Elite Gen 5 QRD | Snapdragon® 8 Elite Gen5 Mobile | ONNX | 3.155 ms | 0 - 20 MB | NPU | SINet.onnx.zip |
| SINet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 1.487 ms | 0 - 0 MB | NPU | SINet.dlc |
| SINet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 6.223 ms | 6 - 6 MB | NPU | SINet.onnx.zip |
Installation
Install the package via pip:
pip install qai-hub-models
Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
Sign-in to Qualcomm® AI Hub 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.
qai-hub configure --api_token API_TOKEN
Navigate to 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.
python -m qai_hub_models.models.sinet.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.sinet.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.
python -m qai_hub_models.models.sinet.export
How does this work?
This export script leverages Qualcomm® AI Hub 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.
import torch
import qai_hub as hub
from qai_hub_models.models.sinet 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.
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.
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.
Run demo on a cloud-hosted device
You can also run the demo on-device.
python -m qai_hub_models.models.sinet.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.sinet.demo -- --eval-mode on-device
Deploying compiled model to Android
The models can be deployed using multiple runtimes:
TensorFlow Lite (
.tfliteexport): This tutorial provides a guide to deploy the .tflite model in an Android application.QNN (
.soexport ): This sample app provides instructions on how to use the.soshared library in an Android application.
View on Qualcomm® AI Hub
Get more details on SINet's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of SINet can be found here.
- The license for the compiled assets for on-device deployment can be found here
References
- SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder
- 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.
- Downloads last month
- 179