Segformer-Base: Optimized for Mobile Deployment

Real-time object segmentation

Segformer Base is a machine learning model that predicts masks and classes of objects in an image.

This model is an implementation of Segformer-Base found here.

This repository provides scripts to run Segformer-Base 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: nvidia/segformer-b0-finetuned-ade-512-512
- Input resolution: 512x512
- Number of output classes: 150
- Number of parameters: 3.75M
- Model size (float): 14.4 MB
- Model size (w8a16): 4.57 MB
- Model size (w8a8): 3.90 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
Segformer-Base	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	221.556 ms	9 - 66 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	221.863 ms	3 - 50 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	114.26 ms	9 - 70 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	118.529 ms	3 - 59 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	114.76 ms	0 - 39 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	113.569 ms	3 - 23 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	117.148 ms	10 - 65 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	116.315 ms	2 - 51 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	221.556 ms	9 - 66 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	221.863 ms	3 - 50 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	114.524 ms	0 - 34 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	114.118 ms	3 - 23 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	125.923 ms	9 - 62 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	124.125 ms	0 - 56 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	114.745 ms	0 - 37 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	113.858 ms	3 - 21 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	117.148 ms	10 - 65 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	116.315 ms	2 - 51 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	114.914 ms	0 - 42 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	114.277 ms	3 - 19 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	110.684 ms	22 - 53 MB	NPU	Segformer-Base.onnx
Segformer-Base	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	86.645 ms	8 - 72 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	86.093 ms	3 - 60 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	84.246 ms	26 - 84 MB	NPU	Segformer-Base.onnx
Segformer-Base	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	76.112 ms	9 - 62 MB	NPU	Segformer-Base.tflite
Segformer-Base	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	92.911 ms	3 - 54 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	94.634 ms	26 - 73 MB	NPU	Segformer-Base.onnx
Segformer-Base	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	119.744 ms	31 - 31 MB	NPU	Segformer-Base.dlc
Segformer-Base	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	118.293 ms	35 - 35 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a16	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	22.12 ms	2 - 37 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	14.453 ms	2 - 46 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	12.227 ms	2 - 14 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	12.605 ms	2 - 38 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	44.72 ms	2 - 69 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	22.12 ms	2 - 37 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	12.347 ms	2 - 14 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	14.846 ms	2 - 38 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	12.255 ms	2 - 13 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	12.605 ms	2 - 38 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	12.298 ms	2 - 14 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	39.84 ms	31 - 105 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	8.226 ms	2 - 51 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	27.765 ms	24 - 282 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	7.417 ms	2 - 43 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	29.544 ms	29 - 241 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	13.624 ms	1 - 1 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	45.624 ms	58 - 58 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	23.299 ms	2 - 40 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	12.961 ms	1 - 33 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	13.628 ms	2 - 47 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	8.276 ms	1 - 43 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	12.945 ms	2 - 12 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	7.134 ms	0 - 10 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	13.523 ms	2 - 41 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	7.764 ms	1 - 34 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	TFLITE	126.846 ms	15 - 55 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	26.472 ms	1 - 49 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	TFLITE	425.541 ms	45 - 61 MB	GPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA7255P ADP	Qualcomm® SA7255P	TFLITE	23.299 ms	2 - 40 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	12.961 ms	1 - 33 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	13.006 ms	2 - 17 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	7.105 ms	0 - 10 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	SA8295P ADP	Qualcomm® SA8295P	TFLITE	15.496 ms	2 - 42 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	8.945 ms	1 - 35 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	12.996 ms	2 - 17 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	7.109 ms	1 - 8 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	SA8775P ADP	Qualcomm® SA8775P	TFLITE	13.523 ms	2 - 41 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	7.764 ms	1 - 34 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	12.983 ms	2 - 13 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	7.108 ms	1 - 11 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	17.99 ms	3 - 25 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	8.881 ms	2 - 49 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	4.766 ms	1 - 45 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	12.54 ms	7 - 71 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	9.246 ms	2 - 44 MB	NPU	Segformer-Base.tflite
Segformer-Base	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	4.355 ms	1 - 40 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	11.89 ms	7 - 59 MB	NPU	Segformer-Base.onnx
Segformer-Base	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	7.993 ms	0 - 0 MB	NPU	Segformer-Base.dlc
Segformer-Base	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	19.401 ms	9 - 9 MB	NPU	Segformer-Base.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.segformer_base.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.segformer_base.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.segformer_base.export

Profiling Results
------------------------------------------------------------
Segformer-Base
Device                          : cs_8275 (ANDROID 14)                 
Runtime                         : TFLITE                               
Estimated inference time (ms)   : 221.6                                
Estimated peak memory usage (MB): [9, 66]                              
Total # Ops                     : 545                                  
Compute Unit(s)                 : npu (545 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.segformer_base 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.segformer_base.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.segformer_base.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 Segformer-Base's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of Segformer-Base can be found here.
The license for the compiled assets for on-device deployment 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.