ResNeXt101: Optimized for Mobile Deployment

Imagenet classifier and general purpose backbone

ResNeXt101 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 ResNeXt101 found here.

This repository provides scripts to run ResNeXt101 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: 88.7M
- Model size (float): 338 MB
- Model size (w8a8): 87.3 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
ResNeXt101	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	36.562 ms	15 - 331 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	35.946 ms	1 - 129 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	9.284 ms	0 - 286 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	9.875 ms	1 - 79 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	6.59 ms	0 - 36 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	7.076 ms	0 - 43 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	10.462 ms	0 - 317 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	10.314 ms	1 - 130 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	36.562 ms	15 - 331 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	35.946 ms	1 - 129 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	6.888 ms	0 - 34 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	6.966 ms	0 - 44 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	10.867 ms	0 - 285 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	10.744 ms	1 - 74 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	6.662 ms	0 - 35 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	7.103 ms	1 - 42 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	10.462 ms	0 - 317 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	10.314 ms	1 - 130 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	6.643 ms	0 - 36 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	7.204 ms	0 - 42 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	6.883 ms	0 - 403 MB	NPU	ResNeXt101.onnx
ResNeXt101	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	4.736 ms	0 - 317 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	4.935 ms	1 - 127 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	4.828 ms	0 - 116 MB	NPU	ResNeXt101.onnx
ResNeXt101	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	4.632 ms	0 - 312 MB	NPU	ResNeXt101.tflite
ResNeXt101	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	4.549 ms	1 - 118 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	5.03 ms	1 - 134 MB	NPU	ResNeXt101.onnx
ResNeXt101	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	7.063 ms	594 - 594 MB	NPU	ResNeXt101.dlc
ResNeXt101	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	6.557 ms	173 - 173 MB	NPU	ResNeXt101.onnx
ResNeXt101	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	6.349 ms	0 - 159 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	6.598 ms	0 - 143 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	3.278 ms	0 - 173 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	4.356 ms	0 - 159 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	2.801 ms	0 - 42 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	2.904 ms	0 - 37 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	3.32 ms	0 - 158 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	3.387 ms	0 - 143 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	15.373 ms	0 - 213 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	TFLITE	131.815 ms	0 - 244 MB	GPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA7255P ADP	Qualcomm® SA7255P	TFLITE	6.349 ms	0 - 159 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	6.598 ms	0 - 143 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	2.772 ms	0 - 39 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	2.943 ms	0 - 37 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	SA8295P ADP	Qualcomm® SA8295P	TFLITE	4.076 ms	0 - 159 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	4.149 ms	0 - 146 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	2.778 ms	0 - 49 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	2.947 ms	0 - 36 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	SA8775P ADP	Qualcomm® SA8775P	TFLITE	3.32 ms	0 - 158 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	3.387 ms	0 - 143 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	2.789 ms	0 - 45 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	2.967 ms	0 - 36 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	41.928 ms	4 - 270 MB	NPU	ResNeXt101.onnx
ResNeXt101	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	2.09 ms	0 - 170 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	2.165 ms	0 - 156 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	27.81 ms	4 - 1016 MB	NPU	ResNeXt101.onnx
ResNeXt101	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	2.043 ms	0 - 161 MB	NPU	ResNeXt101.tflite
ResNeXt101	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	2.129 ms	0 - 147 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	30.626 ms	4 - 1094 MB	NPU	ResNeXt101.onnx
ResNeXt101	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	3.083 ms	360 - 360 MB	NPU	ResNeXt101.dlc
ResNeXt101	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	41.702 ms	107 - 107 MB	NPU	ResNeXt101.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.resnext101.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.resnext101.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.resnext101.export

Profiling Results
------------------------------------------------------------
ResNeXt101
Device                          : cs_8275 (ANDROID 14)                 
Runtime                         : TFLITE                               
Estimated inference time (ms)   : 36.6                                 
Estimated peak memory usage (MB): [15, 331]                            
Total # Ops                     : 147                                  
Compute Unit(s)                 : npu (147 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.resnext101 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.resnext101.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.resnext101.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 ResNeXt101's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of ResNeXt101 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.