RegNet: Optimized for Mobile Deployment

Imagenet classifier and general purpose backbone

RegNet 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 RegNet found here.

This repository provides scripts to run RegNet 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: 15.3M
    • Model size (float): 58.3 MB
    • Model size (w8a8): 15.4 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
RegNet float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 9.883 ms 0 - 76 MB NPU RegNet.tflite
RegNet float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 9.889 ms 1 - 44 MB NPU RegNet.dlc
RegNet float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 2.793 ms 0 - 80 MB NPU RegNet.tflite
RegNet float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 3.319 ms 1 - 45 MB NPU RegNet.dlc
RegNet float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 2.003 ms 0 - 225 MB NPU RegNet.tflite
RegNet float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 1.953 ms 0 - 14 MB NPU RegNet.dlc
RegNet float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 3.249 ms 0 - 77 MB NPU RegNet.tflite
RegNet float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 3.234 ms 0 - 43 MB NPU RegNet.dlc
RegNet float SA7255P ADP Qualcomm® SA7255P TFLITE 9.883 ms 0 - 76 MB NPU RegNet.tflite
RegNet float SA7255P ADP Qualcomm® SA7255P QNN_DLC 9.889 ms 1 - 44 MB NPU RegNet.dlc
RegNet float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 2.0 ms 0 - 225 MB NPU RegNet.tflite
RegNet float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 1.974 ms 1 - 15 MB NPU RegNet.dlc
RegNet float SA8295P ADP Qualcomm® SA8295P TFLITE 3.434 ms 0 - 67 MB NPU RegNet.tflite
RegNet float SA8295P ADP Qualcomm® SA8295P QNN_DLC 3.36 ms 1 - 39 MB NPU RegNet.dlc
RegNet float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 1.999 ms 0 - 229 MB NPU RegNet.tflite
RegNet float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 1.977 ms 1 - 14 MB NPU RegNet.dlc
RegNet float SA8775P ADP Qualcomm® SA8775P TFLITE 3.249 ms 0 - 77 MB NPU RegNet.tflite
RegNet float SA8775P ADP Qualcomm® SA8775P QNN_DLC 3.234 ms 0 - 43 MB NPU RegNet.dlc
RegNet float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 2.008 ms 0 - 222 MB NPU RegNet.tflite
RegNet float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 1.968 ms 1 - 15 MB NPU RegNet.dlc
RegNet float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 1.971 ms 0 - 72 MB NPU RegNet.onnx
RegNet float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 1.403 ms 0 - 91 MB NPU RegNet.tflite
RegNet float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 1.421 ms 1 - 53 MB NPU RegNet.dlc
RegNet float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 1.424 ms 0 - 58 MB NPU RegNet.onnx
RegNet float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 1.379 ms 0 - 79 MB NPU RegNet.tflite
RegNet float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 1.307 ms 0 - 48 MB NPU RegNet.dlc
RegNet float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 1.382 ms 1 - 49 MB NPU RegNet.onnx
RegNet float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 2.657 ms 121 - 121 MB NPU RegNet.dlc
RegNet float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 2.013 ms 39 - 39 MB NPU RegNet.onnx
RegNet w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 2.028 ms 0 - 48 MB NPU RegNet.tflite
RegNet w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 2.285 ms 0 - 46 MB NPU RegNet.dlc
RegNet w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 0.962 ms 0 - 74 MB NPU RegNet.tflite
RegNet w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 1.339 ms 0 - 63 MB NPU RegNet.dlc
RegNet w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 0.759 ms 0 - 68 MB NPU RegNet.tflite
RegNet w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 0.94 ms 0 - 39 MB NPU RegNet.dlc
RegNet w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 1.111 ms 0 - 48 MB NPU RegNet.tflite
RegNet w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 1.261 ms 0 - 46 MB NPU RegNet.dlc
RegNet w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) TFLITE 2.495 ms 0 - 55 MB NPU RegNet.tflite
RegNet w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) QNN_DLC 3.934 ms 0 - 54 MB NPU RegNet.dlc
RegNet w8a8 RB5 (Proxy) Qualcomm® QCS8250 (Proxy) TFLITE 51.771 ms 3 - 85 MB GPU RegNet.tflite
RegNet w8a8 SA7255P ADP Qualcomm® SA7255P TFLITE 2.028 ms 0 - 48 MB NPU RegNet.tflite
RegNet w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 2.285 ms 0 - 46 MB NPU RegNet.dlc
RegNet w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 0.765 ms 0 - 69 MB NPU RegNet.tflite
RegNet w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 0.925 ms 0 - 39 MB NPU RegNet.dlc
RegNet w8a8 SA8295P ADP Qualcomm® SA8295P TFLITE 1.366 ms 0 - 47 MB NPU RegNet.tflite
RegNet w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 1.589 ms 0 - 48 MB NPU RegNet.dlc
RegNet w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 0.782 ms 0 - 68 MB NPU RegNet.tflite
RegNet w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 0.916 ms 0 - 48 MB NPU RegNet.dlc
RegNet w8a8 SA8775P ADP Qualcomm® SA8775P TFLITE 1.111 ms 0 - 48 MB NPU RegNet.tflite
RegNet w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 1.261 ms 0 - 46 MB NPU RegNet.dlc
RegNet w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 0.763 ms 0 - 68 MB NPU RegNet.tflite
RegNet w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 0.916 ms 0 - 29 MB NPU RegNet.dlc
RegNet w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 8.574 ms 0 - 101 MB NPU RegNet.onnx
RegNet w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 0.556 ms 0 - 72 MB NPU RegNet.tflite
RegNet w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 0.644 ms 0 - 67 MB NPU RegNet.dlc
RegNet w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 7.024 ms 4 - 361 MB NPU RegNet.onnx
RegNet w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 0.53 ms 0 - 52 MB NPU RegNet.tflite
RegNet w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 0.585 ms 0 - 51 MB NPU RegNet.dlc
RegNet w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 6.623 ms 2 - 218 MB NPU RegNet.onnx
RegNet w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 1.09 ms 84 - 84 MB NPU RegNet.dlc
RegNet w8a8 Snapdragon X Elite CRD Snapdragon® X Elite ONNX 10.085 ms 22 - 22 MB NPU RegNet.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.regnet.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.regnet.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.regnet.export
Profiling Results
------------------------------------------------------------
RegNet
Device                          : cs_8275 (ANDROID 14)                 
Runtime                         : TFLITE                               
Estimated inference time (ms)   : 9.9                                  
Estimated peak memory usage (MB): [0, 76]                              
Total # Ops                     : 114                                  
Compute Unit(s)                 : npu (114 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.regnet 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.regnet.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.regnet.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 RegNet's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

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

References

Community

Downloads last month
131
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support