Hugging Face's logo

qualcomm
/
ResNet-2Plus1D

Video Classification
PyTorch
LiteRT
ONNX
backbone
android
Model card Files Community

ResNet-2Plus1D: Optimized for Mobile Deployment

Sports and human action recognition in videos

ResNet (2+1)D Convolutions is a network which explicitly factorizes 3D convolution into two separate and successive operations, a 2D spatial convolution and a 1D temporal convolution. It used for video understanding applications.

This model is an implementation of ResNet-2Plus1D found here.

This repository provides scripts to run ResNet-2Plus1D on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

  • Model Type: Model_use_case.video_classification
  • Model Stats:
    • Model checkpoint: Kinetics-400
    • Input resolution: 112x112
    • Number of parameters: 31.5M
    • Model size (float): 120 MB
    • Model size (w8a8): 30.8 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
ResNet-2Plus1D float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 281.969 ms 28 - 66 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 82.344 ms 0 - 65 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 121.99 ms 28 - 94 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 25.597 ms 2 - 85 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 113.792 ms 3 - 746 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 12.791 ms 2 - 28 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 126.567 ms 28 - 67 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 21.594 ms 1 - 64 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float SA7255P ADP Qualcomm® SA7255P TFLITE 281.969 ms 28 - 66 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float SA7255P ADP Qualcomm® SA7255P QNN_DLC 82.344 ms 0 - 65 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 114.514 ms 7 - 789 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 12.816 ms 2 - 21 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float SA8295P ADP Qualcomm® SA8295P TFLITE 139.695 ms 28 - 65 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float SA8295P ADP Qualcomm® SA8295P QNN_DLC 23.162 ms 0 - 55 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 115.413 ms 7 - 766 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 12.802 ms 2 - 25 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float SA8775P ADP Qualcomm® SA8775P TFLITE 126.567 ms 28 - 67 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float SA8775P ADP Qualcomm® SA8775P QNN_DLC 21.594 ms 1 - 64 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 114.486 ms 0 - 768 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 12.778 ms 2 - 25 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 12.083 ms 0 - 146 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 86.573 ms 26 - 86 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 9.271 ms 2 - 91 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 9.346 ms 2 - 96 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 100.682 ms 26 - 66 MB NPU ResNet-2Plus1D.tflite
ResNet-2Plus1D float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 9.006 ms 2 - 70 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 8.909 ms 2 - 70 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 13.799 ms 867 - 867 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 12.68 ms 61 - 61 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 12.232 ms 1 - 39 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 5.83 ms 1 - 62 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 3.927 ms 0 - 12 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 4.334 ms 1 - 40 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) QNN_DLC 24.098 ms 1 - 49 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 12.232 ms 1 - 39 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 3.912 ms 1 - 12 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 7.033 ms 1 - 41 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 3.926 ms 1 - 11 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 4.334 ms 1 - 40 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 3.929 ms 1 - 17 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 3.768 ms 0 - 68 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 2.829 ms 1 - 59 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 2.622 ms 0 - 59 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 2.54 ms 1 - 46 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 2.654 ms 0 - 44 MB NPU ResNet-2Plus1D.onnx
ResNet-2Plus1D w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 4.475 ms 310 - 310 MB NPU ResNet-2Plus1D.dlc
ResNet-2Plus1D w8a8 Snapdragon X Elite CRD Snapdragon® X Elite ONNX 3.731 ms 31 - 31 MB NPU ResNet-2Plus1D.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[resnet-2plus1d]"

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.resnet_2plus1d.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.resnet_2plus1d.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.resnet_2plus1d.export

Profiling Results
------------------------------------------------------------
ResNet-2Plus1D
Device                          : cs_8275 (ANDROID 14)                
Runtime                         : TFLITE                              
Estimated inference time (ms)   : 282.0                               
Estimated peak memory usage (MB): [28, 66]                            
Total # Ops                     : 94                                  
Compute Unit(s)                 : npu (87 ops) gpu (0 ops) cpu (7 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.resnet_2plus1d 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.

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 ResNet-2Plus1D's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of ResNet-2Plus1D can be found here.
  • The license for the compiled assets for on-device deployment can be found here

References

Community

Downloads last month
19
Inference Providers NEW
Video Classification
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support