ResNet-3D: Optimized for Mobile Deployment

Sports and human action recognition in videos

ResNet 3D is a network with 3D convolutions used for video understanding.

This model is an implementation of ResNet-3D found here.

This repository provides scripts to run ResNet-3D 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: 33.4M
  • Model size (float): 127 MB
  • Model size (w8a8): 32.1 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
ResNet-3D	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	571.504 ms	0 - 184 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	91.333 ms	1 - 172 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	324.437 ms	0 - 238 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	26.414 ms	2 - 216 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	303.707 ms	0 - 3 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	13.463 ms	2 - 5 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	13.33 ms	0 - 80 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	1317.716 ms	0 - 185 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	24.215 ms	2 - 193 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	571.504 ms	0 - 184 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	91.333 ms	1 - 172 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	302.322 ms	0 - 2 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	13.466 ms	2 - 4 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	329.574 ms	0 - 182 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	25.789 ms	2 - 178 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	303.687 ms	0 - 3 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	13.465 ms	2 - 5 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	1317.716 ms	0 - 185 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	24.215 ms	2 - 193 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	217.844 ms	0 - 248 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	9.462 ms	2 - 238 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	9.675 ms	2 - 189 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	TFLITE	210.267 ms	0 - 186 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_DLC	7.611 ms	0 - 180 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	7.991 ms	1 - 130 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	TFLITE	190.76 ms	0 - 184 MB	NPU	ResNet-3D.tflite
ResNet-3D	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_DLC	5.848 ms	2 - 186 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	6.004 ms	2 - 135 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	13.895 ms	2 - 2 MB	NPU	ResNet-3D.dlc
ResNet-3D	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	13.527 ms	64 - 64 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	TFLITE	1515.831 ms	678 - 812 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	QNN_DLC	79.864 ms	1 - 161 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Dragonwing Q-6690 MTP	Qualcomm® Qcm6690	ONNX	330.263 ms	59 - 72 MB	CPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	TFLITE	1740.677 ms	695 - 1060 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	QNN_DLC	18.3 ms	1 - 3 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Dragonwing RB3 Gen 2 Vision Kit	Qualcomm® QCS6490	ONNX	266.615 ms	56 - 128 MB	CPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	479.474 ms	0 - 211 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	14.692 ms	1 - 147 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	248.814 ms	0 - 275 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	6.255 ms	1 - 204 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	225.312 ms	0 - 3 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	4.45 ms	1 - 3 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	ONNX	4.773 ms	0 - 42 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	233.199 ms	0 - 209 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	4.558 ms	1 - 147 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	TFLITE	1359.596 ms	1104 - 1127 MB	CPU	ResNet-3D.tflite
ResNet-3D	w8a8	RB5 (Proxy)	Qualcomm® QCS8250 (Proxy)	ONNX	272.868 ms	52 - 96 MB	CPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	SA7255P ADP	Qualcomm® SA7255P	TFLITE	479.474 ms	0 - 211 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	14.692 ms	1 - 147 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	226.926 ms	0 - 3 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	4.469 ms	1 - 2 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	SA8295P ADP	Qualcomm® SA8295P	TFLITE	255.582 ms	0 - 269 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	8.195 ms	1 - 152 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	231.244 ms	0 - 2 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	4.457 ms	1 - 2 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	SA8775P ADP	Qualcomm® SA8775P	TFLITE	233.199 ms	0 - 209 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	4.558 ms	1 - 147 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	172.818 ms	0 - 282 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	3.305 ms	1 - 215 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	3.414 ms	0 - 190 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	TFLITE	127.751 ms	0 - 217 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	QNN_DLC	2.583 ms	1 - 150 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Samsung Galaxy S25	Snapdragon® 8 Elite For Galaxy Mobile	ONNX	2.824 ms	0 - 123 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	TFLITE	1184.654 ms	679 - 841 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	QNN_DLC	7.52 ms	1 - 143 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Snapdragon 7 Gen 4 QRD	Snapdragon® 7 Gen 4 Mobile	ONNX	305.592 ms	43 - 58 MB	CPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	TFLITE	148.393 ms	0 - 257 MB	NPU	ResNet-3D.tflite
ResNet-3D	w8a8	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	QNN_DLC	1.837 ms	1 - 154 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Snapdragon 8 Elite Gen 5 QRD	Snapdragon® 8 Elite Gen5 Mobile	ONNX	2.134 ms	1 - 119 MB	NPU	ResNet-3D.onnx.zip
ResNet-3D	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	4.753 ms	1 - 1 MB	NPU	ResNet-3D.dlc
ResNet-3D	w8a8	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	4.778 ms	33 - 33 MB	NPU	ResNet-3D.onnx.zip

Installation

Install the package via pip:

# NOTE: 3.10 <= PYTHON_VERSION < 3.14 is supported.
pip install "qai-hub-models[resnet-3d]"

Configure Qualcomm® AI Hub Workbench to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub Workbench 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_3d.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_3d.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_3d.export

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_3d import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S25")

# 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 Workbench. 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-3D's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

• The license for the original implementation of ResNet-3D can be found here.

References

• A Closer Look at Spatiotemporal Convolutions for Action Recognition
• Source Model Implementation

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.