ResNet-Mixed-Convolution: Optimized for Mobile Deployment

Sports and human action recognition in videos

ResNet Mixed Convolutions is a network with a mixture of 2D and 3D convolutions used for video understanding.

This model is an implementation of ResNet-Mixed-Convolution found here.

This repository provides scripts to run ResNet-Mixed-Convolution 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: Kinectics-400
- Input resolution: 112x112
- Number of parameters: 11.7M
- Model size (float): 44.7 MB
- Model size (w8a8): 11.7 MB

Model	Precision	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit	Target Model
ResNet-Mixed-Convolution	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	TFLITE	337.464 ms	31 - 81 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	98.496 ms	2 - 72 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	TFLITE	145.438 ms	31 - 70 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	27.864 ms	1 - 57 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	TFLITE	135.337 ms	14 - 78 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	14.006 ms	2 - 25 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	TFLITE	150.633 ms	31 - 81 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	25.775 ms	0 - 75 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	SA7255P ADP	Qualcomm® SA7255P	TFLITE	337.464 ms	31 - 81 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	98.496 ms	2 - 72 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	TFLITE	142.205 ms	31 - 74 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	14.098 ms	2 - 27 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	SA8295P ADP	Qualcomm® SA8295P	TFLITE	163.17 ms	31 - 62 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	27.623 ms	0 - 51 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	TFLITE	136.674 ms	31 - 41 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	14.062 ms	2 - 26 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	SA8775P ADP	Qualcomm® SA8775P	TFLITE	150.633 ms	31 - 81 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	25.775 ms	0 - 75 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	TFLITE	139.769 ms	31 - 92 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	14.085 ms	2 - 28 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	13.648 ms	0 - 71 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	TFLITE	103.846 ms	31 - 89 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	10.025 ms	0 - 85 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	10.412 ms	2 - 90 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	TFLITE	126.007 ms	30 - 79 MB	NPU	ResNet-Mixed-Convolution.tflite
ResNet-Mixed-Convolution	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	10.72 ms	2 - 79 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	10.023 ms	2 - 83 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	float	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	16.103 ms	291 - 291 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	float	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	15.045 ms	22 - 22 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	w8a16	QCS8275 (Proxy)	Qualcomm® QCS8275 (Proxy)	QNN_DLC	28.01 ms	1 - 48 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	QCS8450 (Proxy)	Qualcomm® QCS8450 (Proxy)	QNN_DLC	12.751 ms	1 - 67 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	QCS8550 (Proxy)	Qualcomm® QCS8550 (Proxy)	QNN_DLC	8.355 ms	1 - 19 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	QCS9075 (Proxy)	Qualcomm® QCS9075 (Proxy)	QNN_DLC	8.892 ms	1 - 52 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	RB3 Gen 2 (Proxy)	Qualcomm® QCS6490 (Proxy)	QNN_DLC	54.711 ms	0 - 53 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	SA7255P ADP	Qualcomm® SA7255P	QNN_DLC	28.01 ms	1 - 48 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	SA8255 (Proxy)	Qualcomm® SA8255P (Proxy)	QNN_DLC	8.383 ms	1 - 18 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	SA8295P ADP	Qualcomm® SA8295P	QNN_DLC	14.741 ms	1 - 50 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	SA8650 (Proxy)	Qualcomm® SA8650P (Proxy)	QNN_DLC	8.341 ms	1 - 18 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	SA8775P ADP	Qualcomm® SA8775P	QNN_DLC	8.892 ms	1 - 52 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	QNN_DLC	8.37 ms	1 - 17 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	Samsung Galaxy S23	Snapdragon® 8 Gen 2 Mobile	ONNX	7.78 ms	0 - 31 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	QNN_DLC	6.263 ms	1 - 63 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	Samsung Galaxy S24	Snapdragon® 8 Gen 3 Mobile	ONNX	5.904 ms	0 - 62 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	QNN_DLC	5.839 ms	1 - 48 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite Mobile	ONNX	5.708 ms	1 - 48 MB	NPU	ResNet-Mixed-Convolution.onnx
ResNet-Mixed-Convolution	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN_DLC	9.986 ms	29 - 29 MB	NPU	ResNet-Mixed-Convolution.dlc
ResNet-Mixed-Convolution	w8a16	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	8.298 ms	11 - 11 MB	NPU	ResNet-Mixed-Convolution.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[resnet-mixed]"

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_mixed.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_mixed.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_mixed.export

Profiling Results
------------------------------------------------------------
ResNet-Mixed-Convolution
Device                          : cs_8275 (ANDROID 14)                
Runtime                         : TFLITE                              
Estimated inference time (ms)   : 337.5                               
Estimated peak memory usage (MB): [31, 81]                            
Total # Ops                     : 57                                  
Compute Unit(s)                 : npu (53 ops) gpu (0 ops) cpu (4 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_mixed 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-Mixed-Convolution's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of ResNet-Mixed-Convolution 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.