Hugging Face's logo

qualcomm
/
DDRNet23-Slim

Image Segmentation
PyTorch
LiteRT
real_time
android
Model card Files
xet
Community

DDRNet23-Slim: Optimized for Mobile Deployment

Segment images or video by class in real-time on device

DDRNet23Slim is a machine learning model that segments an image into semantic classes, specifically designed for road-based scenes. It is designed for the application of self-driving cars.

This model is an implementation of DDRNet23-Slim found here.

This repository provides scripts to run DDRNet23-Slim 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: DDRNet23s_imagenet.pth
    • Inference latency: RealTime
    • Input resolution: 2048x1024
    • Number of output classes: 19
    • Number of parameters: 6.13M
    • Model size (float): 21.7 MB
    • Model size (w8a8): 6.11 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
DDRNet23-Slim float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 94.17 ms 2 - 175 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 94.127 ms 24 - 192 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 59.585 ms 2 - 265 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 60.235 ms 24 - 276 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 30.497 ms 0 - 2 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 30.449 ms 24 - 26 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 24.761 ms 24 - 41 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 37.971 ms 2 - 175 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 171.425 ms 21 - 188 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float SA7255P ADP Qualcomm® SA7255P TFLITE 94.17 ms 2 - 175 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float SA7255P ADP Qualcomm® SA7255P QNN_DLC 94.127 ms 24 - 192 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 30.469 ms 2 - 6 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 30.645 ms 24 - 27 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float SA8295P ADP Qualcomm® SA8295P TFLITE 42.222 ms 2 - 184 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float SA8295P ADP Qualcomm® SA8295P QNN_DLC 42.185 ms 24 - 198 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 30.652 ms 2 - 6 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 30.344 ms 24 - 26 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float SA8775P ADP Qualcomm® SA8775P TFLITE 37.971 ms 2 - 175 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float SA8775P ADP Qualcomm® SA8775P QNN_DLC 171.425 ms 21 - 188 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 20.627 ms 1 - 260 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 20.99 ms 17 - 270 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 16.159 ms 32 - 252 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 15.271 ms 1 - 191 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 15.226 ms 16 - 198 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 12.691 ms 5 - 146 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile TFLITE 11.059 ms 2 - 201 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 10.994 ms 24 - 218 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile ONNX 8.955 ms 29 - 218 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 31.403 ms 24 - 24 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 24.375 ms 24 - 24 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 TFLITE 214.038 ms 10 - 198 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Dragonwing Q-6690 MTP Qualcomm® Qcm6690 ONNX 268.402 ms 201 - 218 MB CPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Dragonwing RB3 Gen 2 Vision Kit Qualcomm® QCS6490 TFLITE 180.82 ms 9 - 77 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Dragonwing RB3 Gen 2 Vision Kit Qualcomm® QCS6490 ONNX 299.07 ms 198 - 217 MB CPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 128.224 ms 1 - 169 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 141.159 ms 6 - 175 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 72.963 ms 1 - 221 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 78.492 ms 6 - 225 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 51.185 ms 1 - 3 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 73.592 ms 6 - 8 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) ONNX 74.892 ms 81 - 92 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 51.871 ms 1 - 169 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 74.508 ms 6 - 175 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 RB5 (Proxy) Qualcomm® QCS8250 (Proxy) TFLITE 156.739 ms 24 - 56 MB GPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 RB5 (Proxy) Qualcomm® QCS8250 (Proxy) ONNX 222.896 ms 194 - 203 MB CPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 SA7255P ADP Qualcomm® SA7255P TFLITE 128.224 ms 1 - 169 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 141.159 ms 6 - 175 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 51.578 ms 1 - 3 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 73.609 ms 6 - 8 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 SA8295P ADP Qualcomm® SA8295P TFLITE 63.44 ms 1 - 172 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 82.789 ms 6 - 178 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 51.213 ms 1 - 3 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 73.51 ms 6 - 8 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 SA8775P ADP Qualcomm® SA8775P TFLITE 51.871 ms 1 - 169 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 74.508 ms 6 - 175 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 38.618 ms 0 - 219 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 55.142 ms 6 - 224 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 57.72 ms 92 - 286 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile TFLITE 64.995 ms 1 - 182 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile QNN_DLC 40.451 ms 6 - 184 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 Samsung Galaxy S25 Snapdragon® 8 Elite For Galaxy Mobile ONNX 42.233 ms 84 - 227 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile TFLITE 82.361 ms 9 - 178 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Snapdragon 7 Gen 4 QRD Snapdragon® 7 Gen 4 Mobile ONNX 249.96 ms 207 - 224 MB CPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile TFLITE 40.142 ms 3 - 196 MB NPU DDRNet23-Slim.tflite
DDRNet23-Slim w8a8 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile QNN_DLC 47.001 ms 6 - 201 MB NPU DDRNet23-Slim.dlc
DDRNet23-Slim w8a8 Snapdragon 8 Elite Gen 5 QRD Snapdragon® 8 Elite Gen5 Mobile ONNX 41.735 ms 92 - 235 MB NPU DDRNet23-Slim.onnx.zip
DDRNet23-Slim w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 76.973 ms 6 - 6 MB NPU DDRNet23-Slim.dlc

Installation

Install the package via pip:

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 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.ddrnet23_slim.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.ddrnet23_slim.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.ddrnet23_slim.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.ddrnet23_slim 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 Workbench. 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.ddrnet23_slim.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.ddrnet23_slim.demo -- --eval-mode on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

  • TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

  • QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on DDRNet23-Slim's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of DDRNet23-Slim can be found here.

References

Community

Downloads last month
139
Inference Providers NEW
Image Segmentation
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for qualcomm/DDRNet23-Slim