DeepFace Integration Tutorial#

In this tutorial we use the MXA to accelerate a popular face recognition app DeepFace!

Note

This application assumes the MXA driver/runtimes/compilers have been successfully installed.

Install Deepface#

Clone and install DeepFace repository:

git clone https://github.com/serengil/deepface.git
cd deepface
pip install -e .

Next we will download some sample images to use for verification. Here are some sample images of Brad and George <3. Click to download db and unzip.

unzip db.zip

Next lets write a (tiny) script and run face verification! Open run.py and write the following:

from deepface import DeepFace

MODEL_NAME = 'ArcFaceMX'

def verify(img1_path, img2_path):
    # Run face recognition between the two images
    results = DeepFace.verify(img1_path, img2_path, model_name=MODEL_NAME)

    # Print the result
    qualifier = " " if results['verified'] else " do not "
    print(f'{img1_path} and {img2_path}{qualifier}contain the same person')

def main():
    # Set both images to brad.
    verify('db/brad/brad_0.jpg', 'db/brad/brad_1.jpg')

    # Use deepface to verify the two images
    verify('db/george/george_0.jpg', 'db/brad/brad_1.jpg')

main()

Run the script and it will print the following results:

python3 run.py

db/brad/brad_0.jpg and db/brad/brad_1.jpg contain the same person
db/george/george_0.jpg and db/brad/brad_1.jpg do not contain the same person

This script can distiguish Brad from George!

Accelerate#

In the next step we will accelerate face_recognition with the MXA! This involes two steps:

  1. compile a model

  2. integrate the runtime into the application.

Compile a model#

For simplicty, lets export the ArcFace model built into this app and compile it. Download and export the model:

python3 -c "DeepFace.build_model('ArcFace').model.save('ArcFace.h5')"

Compile the model using the CLI NeuralCompiler tool (mix_nc).

mix_nc -v -m ArcFace.h5

The compiler will create ArcFace.dfp to be used with the accelerator.

Integrate Runtime#

Now we have to integrate the MemryX runtime into this application. DeepFace provides a very nice API to easily add new models, we will leverage this to create a new facial recognition ‘Client’ which will target the MXA to perform inference.

Important

We are currently integrating with a non-streaming application that doesn’t rely on a camera feed or data stream. However, for streaming applications, it is strongly recommended to use the AsyncAccl API or Multi-threading. For additional details, refer to the Inference Tutorials.

In deepface/models/facial_recognition/ArcFace.py add the following code

from typing import Any, Union, List, Tuple
import numpy as np
import memryx as mx
class ArcFaceMXClient(FacialRecognition):
    """
    ArcFace running on the MXA
    """

    def __init__(self):
        self.accl = mx.SyncAccl('ArcFace.dfp')
        self.model_name = "ArcFace"
        self.input_shape = (112, 112)
        self.output_shape = 512

    def forward(self, img: np.ndarray) -> List[float]:
        # Reshape inputs and Perform inference
        ifmap = np.squeeze(img)[:,:,None,:]
        outputs = accl.run(img)
        return np.squeeze(outputs).tolist()

Next, lets register this model in the deepface/modules/modeling.py. Find the following lines and add an entry for the MXClient.

models = {
    "facial_recognition": {
        "VGG-Face": VGGFace.VggFaceClient,
        ...
        "ArcFaceMX": ArcFace.ArcFaceMXClient, # Add this line
        ...
    }

Finally, in the run.py script change the MODEL_NAME to use the accelerated ArcFace.

MODEL_NAME = 'ArcFaceMX'

Rerun the script.

python3 run.py

db/brad/brad_0.jpg and db/brad/brad_1.jpg contain the same person
db/george/george_0.jpg and db/brad/brad_1.jpg do not contain the same person

This script can distiguish Brad from George… but faster!

Final Thoughts#

This example shows the ease of integration using the MemryX SDK using simple CLI tools and runtime APIs.

On this page