PidNet: Optimized for Mobile Deployment
Segment images or video by class in real-time on device
PIDNet (Proportional-Integral-Derivative Network) is a real-time semantic segmentation model based on PID controllers
This model is an implementation of PidNet found here.
This repository provides scripts to run PidNet 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: PIDNet_S_Cityscapes_val.pt
- Inference latency: RealTime
- Input resolution: 1024x2048
- Number of output classes: 19
- Number of parameters: 8.06M
- Model size (float): 29.1 MB
- Model size (w8a8): 8.02 MB
| Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |
|---|---|---|---|---|---|---|---|---|
| PidNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 136.955 ms | 0 - 52 MB | NPU | PidNet.tflite |
| PidNet | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 114.326 ms | 24 - 93 MB | NPU | PidNet.dlc |
| PidNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 57.342 ms | 2 - 69 MB | NPU | PidNet.tflite |
| PidNet | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 63.582 ms | 17 - 102 MB | NPU | PidNet.dlc |
| PidNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 47.158 ms | 2 - 20 MB | NPU | PidNet.tflite |
| PidNet | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 36.05 ms | 24 - 58 MB | NPU | PidNet.dlc |
| PidNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 57.395 ms | 0 - 52 MB | NPU | PidNet.tflite |
| PidNet | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 45.437 ms | 24 - 91 MB | NPU | PidNet.dlc |
| PidNet | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 136.955 ms | 0 - 52 MB | NPU | PidNet.tflite |
| PidNet | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 114.326 ms | 24 - 93 MB | NPU | PidNet.dlc |
| PidNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 47.684 ms | 2 - 23 MB | NPU | PidNet.tflite |
| PidNet | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 36.468 ms | 24 - 57 MB | NPU | PidNet.dlc |
| PidNet | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 64.503 ms | 2 - 55 MB | NPU | PidNet.tflite |
| PidNet | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 52.249 ms | 24 - 105 MB | NPU | PidNet.dlc |
| PidNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 47.417 ms | 2 - 21 MB | NPU | PidNet.tflite |
| PidNet | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 36.101 ms | 24 - 61 MB | NPU | PidNet.dlc |
| PidNet | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 57.395 ms | 0 - 52 MB | NPU | PidNet.tflite |
| PidNet | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 45.437 ms | 24 - 91 MB | NPU | PidNet.dlc |
| PidNet | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 47.366 ms | 2 - 19 MB | NPU | PidNet.tflite |
| PidNet | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN_DLC | 36.566 ms | 24 - 59 MB | NPU | PidNet.dlc |
| PidNet | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 30.736 ms | 29 - 88 MB | NPU | PidNet.onnx |
| PidNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 31.649 ms | 2 - 67 MB | NPU | PidNet.tflite |
| PidNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 26.036 ms | 24 - 94 MB | NPU | PidNet.dlc |
| PidNet | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 20.561 ms | 22 - 92 MB | NPU | PidNet.onnx |
| PidNet | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 32.063 ms | 1 - 56 MB | NPU | PidNet.tflite |
| PidNet | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN_DLC | 24.254 ms | 24 - 95 MB | NPU | PidNet.dlc |
| PidNet | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 18.726 ms | 29 - 92 MB | NPU | PidNet.onnx |
| PidNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 38.853 ms | 39 - 39 MB | NPU | PidNet.dlc |
| PidNet | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 34.492 ms | 24 - 24 MB | NPU | PidNet.onnx |
| PidNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 102.856 ms | 0 - 37 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 143.203 ms | 6 - 60 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 52.384 ms | 1 - 52 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 95.467 ms | 6 - 71 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 51.368 ms | 1 - 17 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 73.244 ms | 6 - 30 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 52.179 ms | 1 - 38 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 74.204 ms | 6 - 62 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 179.658 ms | 2 - 81 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 102.856 ms | 0 - 37 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 143.203 ms | 6 - 60 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 51.54 ms | 0 - 21 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 73.107 ms | 6 - 34 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 61.164 ms | 1 - 41 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 85.321 ms | 6 - 62 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 51.425 ms | 0 - 18 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 73.245 ms | 6 - 30 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 52.179 ms | 1 - 38 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 74.204 ms | 6 - 62 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 51.332 ms | 1 - 21 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN_DLC | 73.083 ms | 6 - 33 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 77.667 ms | 88 - 105 MB | NPU | PidNet.onnx |
| PidNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 39.242 ms | 0 - 48 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 55.524 ms | 6 - 73 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 59.685 ms | 81 - 140 MB | NPU | PidNet.onnx |
| PidNet | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 38.992 ms | 0 - 42 MB | NPU | PidNet.tflite |
| PidNet | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN_DLC | 61.574 ms | 6 - 75 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 65.991 ms | 101 - 154 MB | NPU | PidNet.onnx |
| PidNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 78.307 ms | 19 - 19 MB | NPU | PidNet.dlc |
| PidNet | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 69.065 ms | 132 - 132 MB | NPU | PidNet.onnx |
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.pidnet.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.pidnet.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.pidnet.export
Profiling Results
------------------------------------------------------------
PidNet
Device : cs_8275 (ANDROID 14)
Runtime : TFLITE
Estimated inference time (ms) : 137.0
Estimated peak memory usage (MB): [0, 52]
Total # Ops : 169
Compute Unit(s) : npu (169 ops) gpu (0 ops) cpu (0 ops)
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.pidnet 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.
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.pidnet.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.pidnet.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 PidNet's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of PidNet can be found here.
- The license for the compiled assets for on-device deployment can be found here
References
- PIDNet A Real-time Semantic Segmentation Network Inspired from PID Controller Segmentation of Road Scenes
- 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
- 41