MobileSam: Optimized for Mobile Deployment

FASTER SEGMENT ANYTHING: TOWARDS LIGHTWEIGHT SAM FOR MOBILE APPLICATIONS

Transformer based encoder-decoder where prompts specify what to segment in an image thereby allowing segmentation without the need for additional training. The image encoder generates embeddings and the lightweight decoder operates on the embeddings for point and mask based image segmentation.

This model is an implementation of MobileSam found here.

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

Model Details

Model Type: Semantic segmentation
Model Stats:
- Model checkpoint: vit_t
- Input resolution: 720p (720x1280)
- Number of parameters (MobileSamDecoder): 3.876M
- Model size (MobileSamDecoder): 19.6 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
SAMEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	439.913 ms	34 - 60 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	334.816 ms	12 - 89 MB	FP16	NPU	MobileSam.so
SAMEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	417.54 ms	65 - 122 MB	FP16	NPU	MobileSam.onnx
SAMEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	335.191 ms	33 - 158 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	247.716 ms	191 - 694 MB	FP16	NPU	MobileSam.so
SAMEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	313.436 ms	95 - 218 MB	FP16	NPU	MobileSam.onnx
SAMEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	237.753 ms	33 - 166 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	235.822 ms	12 - 519 MB	FP16	NPU	Use Export Script
SAMEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	271.979 ms	78 - 213 MB	FP16	NPU	MobileSam.onnx
SAMEncoder	SA7255P ADP	SA7255P	TFLITE	1302.15 ms	33 - 166 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	SA7255P ADP	SA7255P	QNN	991.612 ms	12 - 21 MB	FP16	NPU	Use Export Script
SAMEncoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	415.155 ms	33 - 59 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	SA8255 (Proxy)	SA8255P Proxy	QNN	266.684 ms	12 - 15 MB	FP16	NPU	Use Export Script
SAMEncoder	SA8295P ADP	SA8295P	TFLITE	585.405 ms	33 - 168 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	SA8295P ADP	SA8295P	QNN	424.958 ms	0 - 18 MB	FP16	NPU	Use Export Script
SAMEncoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	432.024 ms	33 - 58 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	SA8650 (Proxy)	SA8650P Proxy	QNN	268.519 ms	12 - 15 MB	FP16	NPU	Use Export Script
SAMEncoder	SA8775P ADP	SA8775P	TFLITE	494.361 ms	33 - 165 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	SA8775P ADP	SA8775P	QNN	330.266 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMEncoder	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	1302.15 ms	33 - 166 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	QCS8275 (Proxy)	QCS8275 Proxy	QNN	991.612 ms	12 - 21 MB	FP16	NPU	Use Export Script
SAMEncoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	420.702 ms	33 - 60 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	270.321 ms	12 - 15 MB	FP16	NPU	Use Export Script
SAMEncoder	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	494.361 ms	33 - 165 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	QCS9075 (Proxy)	QCS9075 Proxy	QNN	330.266 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMEncoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	597.124 ms	33 - 171 MB	FP16	NPU	MobileSam.tflite
SAMEncoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	503.439 ms	12 - 572 MB	FP16	NPU	Use Export Script
SAMEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	275.036 ms	12 - 12 MB	FP16	NPU	Use Export Script
SAMEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	448.994 ms	130 - 130 MB	FP16	NPU	MobileSam.onnx
SAMDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	7.38 ms	0 - 29 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	6.435 ms	4 - 21 MB	FP16	NPU	MobileSam.so
SAMDecoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	8.887 ms	1 - 62 MB	FP16	NPU	MobileSam.onnx
SAMDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	5.151 ms	0 - 46 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	4.55 ms	4 - 48 MB	FP16	NPU	MobileSam.so
SAMDecoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	6.067 ms	6 - 75 MB	FP16	NPU	MobileSam.onnx
SAMDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	5.041 ms	0 - 44 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	4.357 ms	4 - 42 MB	FP16	NPU	Use Export Script
SAMDecoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	4.566 ms	4 - 62 MB	FP16	NPU	MobileSam.onnx
SAMDecoder	SA7255P ADP	SA7255P	TFLITE	53.054 ms	0 - 40 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	SA7255P ADP	SA7255P	QNN	48.542 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMDecoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	7.371 ms	0 - 26 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	SA8255 (Proxy)	SA8255P Proxy	QNN	6.184 ms	4 - 6 MB	FP16	NPU	Use Export Script
SAMDecoder	SA8295P ADP	SA8295P	TFLITE	9.906 ms	0 - 36 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	SA8295P ADP	SA8295P	QNN	7.432 ms	0 - 17 MB	FP16	NPU	Use Export Script
SAMDecoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	7.351 ms	0 - 25 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	SA8650 (Proxy)	SA8650P Proxy	QNN	6.18 ms	4 - 7 MB	FP16	NPU	Use Export Script
SAMDecoder	SA8775P ADP	SA8775P	TFLITE	10.347 ms	0 - 40 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	SA8775P ADP	SA8775P	QNN	8.85 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMDecoder	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	53.054 ms	0 - 40 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	QCS8275 (Proxy)	QCS8275 Proxy	QNN	48.542 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMDecoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	7.346 ms	0 - 26 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	6.298 ms	4 - 6 MB	FP16	NPU	Use Export Script
SAMDecoder	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	10.347 ms	0 - 40 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	QCS9075 (Proxy)	QCS9075 Proxy	QNN	8.85 ms	1 - 11 MB	FP16	NPU	Use Export Script
SAMDecoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	8.763 ms	0 - 42 MB	FP16	NPU	MobileSam.tflite
SAMDecoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	8.018 ms	4 - 42 MB	FP16	NPU	Use Export Script
SAMDecoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	6.736 ms	4 - 4 MB	FP16	NPU	Use Export Script
SAMDecoder	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	10.021 ms	12 - 12 MB	FP16	NPU	MobileSam.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[mobilesam]"

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.mobilesam.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.mobilesam.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.mobilesam.export

Profiling Results
------------------------------------------------------------
SAMEncoder
Device                          : Samsung Galaxy S23 (13)   
Runtime                         : TFLITE                    
Estimated inference time (ms)   : 439.9                     
Estimated peak memory usage (MB): [34, 60]                  
Total # Ops                     : 592                       
Compute Unit(s)                 : NPU (532 ops) CPU (60 ops)

------------------------------------------------------------
SAMDecoder
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 7.4                    
Estimated peak memory usage (MB): [0, 29]                
Total # Ops                     : 845                    
Compute Unit(s)                 : NPU (845 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.mobilesam 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.mobilesam.demo --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.mobilesam.demo -- --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 MobileSam's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of MobileSam 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.