DeepFakeDetection-torchserve-deployment

In this repository, I present the process of deploying a deepfake detection model via the use of TorchServe.

Environment

Code: Install requirements

pip install -r requirements.txt

Save the model locally

After the model has been trained, whether it be through a notebook or an end-to-end pipeline, you can run the provided code and copy the model into the directory that has been created (./model/efficientnet-b4-deepfake).

Code: Writing setup_config.json for deepfake model

import os
import json

model_name = 'efficientnet-b4'
model_dir_path = os.path.join('./model/', model_name + '-deepfake')
os.mkdir(model_dir_path)

with open(os.path.join(model_dir_path, 'index_to_name.json'), 'w') as file:
    json.dump({'0': 'original', '1': 'fake'}, file)

with open(os.path.join(model_dir_path, 'setup_config.json'), 'w') as file:
    json.dump({
        'model_name': model_name,
        'num_classes': 1
    }, file)

Build torch-model-archiver

Create the archive of the model
You can check torch-model-archiver --help to learn more about its parameters. Make a folder named model_store before you run the code.

Code: Create the torch-model-archive of the deepfake detection model based on efficientnet-b4

torch-model-archiver -f --model-name "efficientnet-b4-deepfake" --export-path ./model_store --version 1.0 \
--serialized-file ./model/efficientnet-b4-deepfake/model.pth --model-file ./model.py --handler ./handler.py \
--extra-files "./model/efficientnet-b4-deepfake/setup_config.json,./model/efficientnet-b4-deepfake/index_to_name.json"

Run the model on torch-server

In order to run the following codes, it is necessary to have docker on your machine. Additionally, it is important to note that the torch-server will deploy all models in the model_store directory by default.

1. If you want to run torch-server on CPU

Code: Build the docker image

docker build -t torchserve-efficientnet .

Code: Start torch-server

docker run --rm -it --volume $PWD/model_store:/home/model-server/model-store -p8080:8080 -p8081:8081 \
torchserve-efficientnet

2. If you want to run torch-server on GPU

Code: Build the docker image

docker build -t torchserve-efficientnet -f Dockerfile_gpu .

Code: Start torch-server

docker run --rm -it --gpus all --runtime=nvidia --volume $PWD/model_store:/home/model-server/model-store -p8080:8080 \
-p8081:8081 torchserve-efficientnet

Code: Start torch-server with a custom configuration

docker run --rm -it --volume $PWD/model_store:/home/model-server/model-store \
--mount type=bind,source=$PWD/config.properties,target=/home/model-server/config.properties -p8080:8080 -p8081:8081 \
torchserve-efficientnet

Test

Code: To see the status of the models

curl --location --request GET 'http://127.0.0.1:8081/models'

Code: We send an original face image to the "efficientnet-b4-deepfake" request.

curl http://127.0.0.1:8080/predictions/efficientnet-b4-deepfake -T ./examples/original.png

Code: We send a fake face image to the "efficientnet-b4-deepfake" request.

curl http://127.0.0.1:8080/predictions/efficientnet-b4-deepfake -T ./examples/fake.png