ResNeXt50: Optimized for Mobile Deployment

Imagenet classifier and general purpose backbone

ResNeXt50 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 ResNeXt50 found here.

This repository provides scripts to run ResNeXt50 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: 25.0M
    • Model size (float): 95.4 MB
    • Model size (w8a8): 24.8 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
ResNeXt50 float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 12.165 ms 0 - 95 MB NPU ResNeXt50.tflite
ResNeXt50 float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 11.881 ms 1 - 43 MB NPU ResNeXt50.dlc
ResNeXt50 float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 3.355 ms 0 - 96 MB NPU ResNeXt50.tflite
ResNeXt50 float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 3.851 ms 1 - 39 MB NPU ResNeXt50.dlc
ResNeXt50 float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 2.515 ms 0 - 305 MB NPU ResNeXt50.tflite
ResNeXt50 float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 2.445 ms 1 - 16 MB NPU ResNeXt50.dlc
ResNeXt50 float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 3.906 ms 0 - 95 MB NPU ResNeXt50.tflite
ResNeXt50 float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 3.758 ms 1 - 44 MB NPU ResNeXt50.dlc
ResNeXt50 float SA7255P ADP Qualcomm® SA7255P TFLITE 12.165 ms 0 - 95 MB NPU ResNeXt50.tflite
ResNeXt50 float SA7255P ADP Qualcomm® SA7255P QNN_DLC 11.881 ms 1 - 43 MB NPU ResNeXt50.dlc
ResNeXt50 float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 2.492 ms 0 - 307 MB NPU ResNeXt50.tflite
ResNeXt50 float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 2.438 ms 1 - 16 MB NPU ResNeXt50.dlc
ResNeXt50 float SA8295P ADP Qualcomm® SA8295P TFLITE 4.014 ms 0 - 87 MB NPU ResNeXt50.tflite
ResNeXt50 float SA8295P ADP Qualcomm® SA8295P QNN_DLC 3.906 ms 1 - 33 MB NPU ResNeXt50.dlc
ResNeXt50 float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 2.494 ms 0 - 305 MB NPU ResNeXt50.tflite
ResNeXt50 float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 2.433 ms 1 - 15 MB NPU ResNeXt50.dlc
ResNeXt50 float SA8775P ADP Qualcomm® SA8775P TFLITE 3.906 ms 0 - 95 MB NPU ResNeXt50.tflite
ResNeXt50 float SA8775P ADP Qualcomm® SA8775P QNN_DLC 3.758 ms 1 - 44 MB NPU ResNeXt50.dlc
ResNeXt50 float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 2.494 ms 0 - 314 MB NPU ResNeXt50.tflite
ResNeXt50 float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 2.434 ms 1 - 16 MB NPU ResNeXt50.dlc
ResNeXt50 float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 2.445 ms 0 - 161 MB NPU ResNeXt50.onnx
ResNeXt50 float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 1.766 ms 0 - 97 MB NPU ResNeXt50.tflite
ResNeXt50 float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 1.774 ms 1 - 53 MB NPU ResNeXt50.dlc
ResNeXt50 float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 1.812 ms 0 - 53 MB NPU ResNeXt50.onnx
ResNeXt50 float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 1.649 ms 0 - 98 MB NPU ResNeXt50.tflite
ResNeXt50 float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 1.443 ms 1 - 49 MB NPU ResNeXt50.dlc
ResNeXt50 float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 1.79 ms 0 - 47 MB NPU ResNeXt50.onnx
ResNeXt50 float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 3.019 ms 209 - 209 MB NPU ResNeXt50.dlc
ResNeXt50 float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 2.419 ms 50 - 50 MB NPU ResNeXt50.onnx
ResNeXt50 w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 2.203 ms 0 - 49 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 2.471 ms 0 - 47 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 1.122 ms 0 - 62 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 1.436 ms 0 - 59 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 0.92 ms 0 - 92 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 1.084 ms 0 - 17 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 1.241 ms 0 - 50 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 1.424 ms 0 - 46 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) TFLITE 3.047 ms 0 - 65 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) QNN_DLC 4.399 ms 0 - 64 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 RB5 (Proxy) Qualcomm® QCS8250 (Proxy) TFLITE 63.08 ms 0 - 119 MB GPU ResNeXt50.tflite
ResNeXt50 w8a8 SA7255P ADP Qualcomm® SA7255P TFLITE 2.203 ms 0 - 49 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 2.471 ms 0 - 47 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 0.918 ms 0 - 94 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 1.079 ms 0 - 18 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 SA8295P ADP Qualcomm® SA8295P TFLITE 1.47 ms 0 - 51 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 1.677 ms 0 - 48 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 0.902 ms 0 - 92 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 1.089 ms 0 - 18 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 SA8775P ADP Qualcomm® SA8775P TFLITE 1.241 ms 0 - 50 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 1.424 ms 0 - 46 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 0.92 ms 0 - 91 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 1.082 ms 0 - 15 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 1.351 ms 0 - 43 MB NPU ResNeXt50.onnx
ResNeXt50 w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 0.679 ms 0 - 60 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 0.807 ms 0 - 57 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 0.976 ms 0 - 71 MB NPU ResNeXt50.onnx
ResNeXt50 w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 0.639 ms 0 - 58 MB NPU ResNeXt50.tflite
ResNeXt50 w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 0.725 ms 0 - 51 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 1.02 ms 0 - 59 MB NPU ResNeXt50.onnx
ResNeXt50 w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 1.235 ms 110 - 110 MB NPU ResNeXt50.dlc
ResNeXt50 w8a8 Snapdragon X Elite CRD Snapdragon® X Elite ONNX 1.333 ms 29 - 29 MB NPU ResNeXt50.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.resnext50.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.resnext50.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.resnext50.export
Profiling Results
------------------------------------------------------------
ResNeXt50
Device                          : cs_8275 (ANDROID 14)                
Runtime                         : TFLITE                              
Estimated inference time (ms)   : 12.2                                
Estimated peak memory usage (MB): [0, 95]                             
Total # Ops                     : 79                                  
Compute Unit(s)                 : npu (79 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.resnext50 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.resnext50.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.resnext50.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 ResNeXt50's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of ResNeXt50 can be found here.
  • The license for the compiled assets for on-device deployment can be found here

References

Community

Downloads last month
102
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support