GKT: Optimized for Mobile Deployment

Construct a bird’s eye view from sensors mounted on a vehicle

Geometry-guided Kernel Transformer is a machine learning model for generating a birds eye view represenation from the sensors(cameras) mounted on a vehicle.

This model is an implementation of GKT found [here](https://github.com/hustvl/GKT/ https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/Qualcomm+AI+Hub+Proprietary+License.pdf).

This repository provides scripts to run GKT on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Model_use_case.driver_assistance
Model Stats:
- Model checkpoint: map_segmentation_gkt_7x1_conv_setting2.ckpt
- Input resolution: 1 x 6 x 3 x 224 x 480
- Number of parameters: 1.18M
- Model size: 4.66 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
GKT	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_CONTEXT_BINARY	178.364 ms	4 - 13 MB	NPU	Use Export Script
GKT	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_CONTEXT_BINARY	204.304 ms	7 - 25 MB	NPU	Use Export Script
GKT	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_CONTEXT_BINARY	105.774 ms	8 - 10 MB	NPU	Use Export Script
GKT	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	PRECOMPILED_QNN_ONNX	80.205 ms	7 - 12 MB	NPU	Use Export Script
GKT	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_CONTEXT_BINARY	107.699 ms	0 - 9 MB	NPU	Use Export Script
GKT	float	SA7255P ADP	Qualcomm® SA7255P	QNN_CONTEXT_BINARY	178.364 ms	4 - 13 MB	NPU	Use Export Script
GKT	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_CONTEXT_BINARY	106.111 ms	8 - 10 MB	NPU	Use Export Script
GKT	float	SA8295P ADP	Qualcomm® SA8295P	QNN_CONTEXT_BINARY	139.075 ms	0 - 17 MB	NPU	Use Export Script
GKT	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_CONTEXT_BINARY	105.282 ms	8 - 10 MB	NPU	Use Export Script
GKT	float	SA8775P ADP	Qualcomm® SA8775P	QNN_CONTEXT_BINARY	107.699 ms	0 - 9 MB	NPU	Use Export Script
GKT	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_CONTEXT_BINARY	76.659 ms	8 - 28 MB	NPU	Use Export Script
GKT	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	PRECOMPILED_QNN_ONNX	55.17 ms	8 - 26 MB	NPU	Use Export Script
GKT	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_CONTEXT_BINARY	67.44 ms	7 - 20 MB	NPU	Use Export Script
GKT	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	PRECOMPILED_QNN_ONNX	50.897 ms	1 - 11 MB	NPU	Use Export Script
GKT	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_CONTEXT_BINARY	56.469 ms	7 - 18 MB	NPU	Use Export Script
GKT	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	PRECOMPILED_QNN_ONNX	40.318 ms	8 - 18 MB	NPU	Use Export Script
GKT	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_CONTEXT_BINARY	99.942 ms	7 - 7 MB	NPU	Use Export Script
GKT	float	Snapdragon X Elite CRD	Snapdragon® X Elite	PRECOMPILED_QNN_ONNX	77.464 ms	7 - 7 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_CONTEXT_BINARY	192.276 ms	4 - 13 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_CONTEXT_BINARY	145.984 ms	4 - 24 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_CONTEXT_BINARY	130.987 ms	4 - 6 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	PRECOMPILED_QNN_ONNX	128.848 ms	4 - 6 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_CONTEXT_BINARY	130.062 ms	1 - 11 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	SA7255P ADP	Qualcomm® SA7255P	QNN_CONTEXT_BINARY	192.276 ms	4 - 13 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_CONTEXT_BINARY	129.938 ms	4 - 6 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	SA8295P ADP	Qualcomm® SA8295P	QNN_CONTEXT_BINARY	153.875 ms	0 - 17 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_CONTEXT_BINARY	130.476 ms	4 - 7 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	SA8775P ADP	Qualcomm® SA8775P	QNN_CONTEXT_BINARY	130.062 ms	1 - 11 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_CONTEXT_BINARY	93.949 ms	4 - 22 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	PRECOMPILED_QNN_ONNX	97.679 ms	4 - 23 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_CONTEXT_BINARY	82.23 ms	4 - 20 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	PRECOMPILED_QNN_ONNX	83.95 ms	0 - 14 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_CONTEXT_BINARY	74.446 ms	4 - 15 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	PRECOMPILED_QNN_ONNX	70.609 ms	4 - 14 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_CONTEXT_BINARY	130.182 ms	4 - 4 MB	NPU	Use Export Script
GKT	w8a16_mixed_fp16	Snapdragon X Elite CRD	Snapdragon® X Elite	PRECOMPILED_QNN_ONNX	129.483 ms	7 - 7 MB	NPU	Use Export Script

Installation

Install the package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install nuscenes-devkit==1.2.0 --no-deps
pip install "qai-hub-models[gkt]"

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.gkt.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.gkt.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.gkt.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.gkt 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.gkt.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.gkt.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 GKT's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of GKT can be found here.

References

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: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Other

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for qualcomm/GKT

Efficient and Robust 2D-to-BEV Representation Learning via Geometry-guided Kernel Transformer

Paper • 2206.04584 • Published Jun 9, 2022