Video-MAE-Quantized: Optimized for Mobile Deployment

Sports and human action recognition in videos

Video MAE (Masked Auto Encoder) is a network for doing video classification that uses the ViT (Vision Transformer) backbone.

This model is an implementation of Video-MAE-Quantized found here.

This repository provides scripts to run Video-MAE-Quantized on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

• Model Type: Video classification
• Model Stats:
  • Model checkpoint: Kinectics-400
  • Input resolution: 224x224
  • Number of parameters: 87.7M
  • Model size: 87.7 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
Video-MAE-Quantized	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	536.044 ms	2 - 64 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	890.755 ms	244 - 509 MB	INT8	NPU	Video-MAE-Quantized.onnx
Video-MAE-Quantized	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	407.325 ms	2 - 354 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	698.819 ms	507 - 871 MB	INT8	NPU	Video-MAE-Quantized.onnx
Video-MAE-Quantized	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	377.508 ms	1 - 362 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	644.066 ms	513 - 883 MB	INT8	NPU	Video-MAE-Quantized.onnx
Video-MAE-Quantized	SA7255P ADP	SA7255P	TFLITE	2228.922 ms	2 - 363 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	SA8255 (Proxy)	SA8255P Proxy	TFLITE	534.996 ms	2 - 65 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	SA8295P ADP	SA8295P	TFLITE	636.231 ms	2 - 497 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	SA8650 (Proxy)	SA8650P Proxy	TFLITE	535.311 ms	2 - 71 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	SA8775P ADP	SA8775P	TFLITE	656.528 ms	2 - 363 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	2228.922 ms	2 - 363 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	537.05 ms	2 - 72 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	656.528 ms	2 - 363 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	574.015 ms	2 - 515 MB	INT8	NPU	Video-MAE-Quantized.tflite
Video-MAE-Quantized	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	927.056 ms	928 - 928 MB	INT8	NPU	Video-MAE-Quantized.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[video-mae-quantized]"

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.video_mae_quantized.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.video_mae_quantized.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.video_mae_quantized.export

Profiling Results
------------------------------------------------------------
Video-MAE-Quantized
Device                          : Samsung Galaxy S23 (13)  
Runtime                         : TFLITE                   
Estimated inference time (ms)   : 536.0                    
Estimated peak memory usage (MB): [2, 64]                  
Total # Ops                     : 612                      
Compute Unit(s)                 : NPU (611 ops) CPU (1 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.video_mae_quantized 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 Video-MAE-Quantized's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

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

References

• Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
• 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.