VIT: Optimized for Mobile Deployment

Imagenet classifier and general purpose backbone

VIT is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This model is an implementation of VIT found here.

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

Model Details

Model Type: Model_use_case.image_classification
Model Stats:
- Model checkpoint: Imagenet
- Input resolution: 224x224
- Number of parameters: 86.6M
- Model size (float): 330 MB
- Model size (w8a16): 86.2 MB
- Model size (w8a8): 83.2 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
VIT	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	35.901 ms	0 - 433 MB	NPU	VIT.tflite
VIT	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	40.119 ms	1 - 424 MB	NPU	VIT.dlc
VIT	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	14.66 ms	0 - 462 MB	NPU	VIT.tflite
VIT	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	19.999 ms	1 - 451 MB	NPU	VIT.dlc
VIT	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	8.092 ms	0 - 3 MB	NPU	VIT.tflite
VIT	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	11.213 ms	1 - 3 MB	NPU	VIT.dlc
VIT	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	13.336 ms	0 - 194 MB	NPU	VIT.onnx.zip
VIT	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	11.041 ms	0 - 433 MB	NPU	VIT.tflite
VIT	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	14.003 ms	1 - 424 MB	NPU	VIT.dlc
VIT	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	35.901 ms	0 - 433 MB	NPU	VIT.tflite
VIT	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	40.119 ms	1 - 424 MB	NPU	VIT.dlc
VIT	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	8.09 ms	0 - 3 MB	NPU	VIT.tflite
VIT	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	11.242 ms	1 - 3 MB	NPU	VIT.dlc
VIT	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	13.58 ms	0 - 425 MB	NPU	VIT.tflite
VIT	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	17.087 ms	0 - 424 MB	NPU	VIT.dlc
VIT	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	8.081 ms	0 - 3 MB	NPU	VIT.tflite
VIT	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	11.239 ms	1 - 3 MB	NPU	VIT.dlc
VIT	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	11.041 ms	0 - 433 MB	NPU	VIT.tflite
VIT	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	14.003 ms	1 - 424 MB	NPU	VIT.dlc
VIT	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	5.952 ms	0 - 473 MB	NPU	VIT.tflite
VIT	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	7.737 ms	1 - 460 MB	NPU	VIT.dlc
VIT	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	9.097 ms	1 - 476 MB	NPU	VIT.onnx.zip
VIT	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	TFLITE	4.087 ms	0 - 429 MB	NPU	VIT.tflite
VIT	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_DLC	5.473 ms	1 - 427 MB	NPU	VIT.dlc
VIT	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	6.313 ms	1 - 421 MB	NPU	VIT.onnx.zip
VIT	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	TFLITE	3.324 ms	0 - 426 MB	NPU	VIT.tflite
VIT	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_DLC	4.234 ms	1 - 421 MB	NPU	VIT.dlc
VIT	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	4.856 ms	1 - 414 MB	NPU	VIT.onnx.zip
VIT	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	11.898 ms	1 - 1 MB	NPU	VIT.dlc
VIT	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	13.877 ms	171 - 171 MB	NPU	VIT.onnx.zip
VIT	w8a16	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	ONNX	620.662 ms	87 - 108 MB	CPU	VIT.onnx.zip
VIT	w8a16	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	ONNX	1117.411 ms	47 - 75 MB	CPU	VIT.onnx.zip
VIT	w8a16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	277.023 ms	55 - 60 MB	NPU	VIT.onnx.zip
VIT	w8a16	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	ONNX	556.57 ms	86 - 100 MB	CPU	VIT.onnx.zip
VIT	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	242.94 ms	59 - 306 MB	NPU	VIT.onnx.zip
VIT	w8a16	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	202.773 ms	53 - 208 MB	NPU	VIT.onnx.zip
VIT	w8a16	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	ONNX	592.98 ms	86 - 106 MB	CPU	VIT.onnx.zip
VIT	w8a16	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	182.106 ms	57 - 217 MB	NPU	VIT.onnx.zip
VIT	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	167.05 ms	71 - 71 MB	NPU	VIT.onnx.zip
VIT	w8a8	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	TFLITE	88.264 ms	2 - 157 MB	NPU	VIT.tflite
VIT	w8a8	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	ONNX	477.555 ms	69 - 92 MB	CPU	VIT.onnx.zip
VIT	w8a8	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	TFLITE	71.779 ms	1 - 99 MB	NPU	VIT.tflite
VIT	w8a8	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	ONNX	873.83 ms	67 - 100 MB	CPU	VIT.onnx.zip
VIT	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	14.406 ms	0 - 165 MB	NPU	VIT.tflite
VIT	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	8.012 ms	0 - 263 MB	NPU	VIT.tflite
VIT	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	6.753 ms	0 - 4 MB	NPU	VIT.tflite
VIT	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	289.68 ms	42 - 77 MB	NPU	VIT.onnx.zip
VIT	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	32.057 ms	0 - 165 MB	NPU	VIT.tflite
VIT	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	ONNX	429.052 ms	61 - 79 MB	CPU	VIT.onnx.zip
VIT	w8a8	SA7255P ADP	Qualcomm® SA7255P	TFLITE	14.406 ms	0 - 165 MB	NPU	VIT.tflite
VIT	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	6.762 ms	0 - 2 MB	NPU	VIT.tflite
VIT	w8a8	SA8295P ADP	Qualcomm® SA8295P	TFLITE	8.944 ms	0 - 168 MB	NPU	VIT.tflite
VIT	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	6.747 ms	0 - 3 MB	NPU	VIT.tflite
VIT	w8a8	SA8775P ADP	Qualcomm® SA8775P	TFLITE	32.057 ms	0 - 165 MB	NPU	VIT.tflite
VIT	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	4.746 ms	0 - 264 MB	NPU	VIT.tflite
VIT	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	229.346 ms	60 - 278 MB	NPU	VIT.onnx.zip
VIT	w8a8	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	TFLITE	3.389 ms	0 - 170 MB	NPU	VIT.tflite
VIT	w8a8	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	199.154 ms	32 - 175 MB	NPU	VIT.onnx.zip
VIT	w8a8	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	TFLITE	20.39 ms	2 - 149 MB	NPU	VIT.tflite
VIT	w8a8	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	ONNX	461.702 ms	38 - 58 MB	CPU	VIT.onnx.zip
VIT	w8a8	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	TFLITE	2.279 ms	0 - 172 MB	NPU	VIT.tflite
VIT	w8a8	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	181.066 ms	56 - 206 MB	NPU	VIT.onnx.zip
VIT	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	139.815 ms	69 - 69 MB	NPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	ONNX	507.94 ms	89 - 113 MB	CPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	ONNX	893.549 ms	37 - 77 MB	CPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	341.516 ms	51 - 83 MB	NPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	ONNX	432.449 ms	84 - 108 MB	CPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	279.907 ms	78 - 328 MB	NPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	245.32 ms	80 - 242 MB	NPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	ONNX	488.346 ms	87 - 110 MB	CPU	VIT.onnx.zip
VIT	w8a8_mixed_int16	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	225.812 ms	75 - 236 MB	NPU	VIT.onnx.zip

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

License

The license for the original implementation of VIT 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: 125

Paper for qualcomm/VIT

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Paper • 2010.11929 • Published Oct 22, 2020 • 15