This folder contains the neceseray code to run inference using the DeepFake Detection (DFDC) solution by team NTech Lab. For more details on the project please visit the GitHub repository.
The following python packages are required:
- albumentations==0.5.2
- numpy==1.18.1
- torchvision==0.5.0
- efficientnet_pytorch==0.6.3
- opencv_python==4.2.0.34
- matplotlib==3.1.3
- torch==1.4.0
- Flask==1.1.2
They can be installed by running pip install -r requirements.txt.
Additionally, opencv requires the following system libraries libglib2.0-0 libsm6 libxext6 libxrender1 libfontconfig1
To start a container that has all the requirements for using the model first build the image by running the following command in this directory:
docker build -t ntech_i .
To use GPUs in the container you will need to install the NVIDIA container toolkit. To run the models, a directory with the pretrained model weights will need to be mounted in the container's workdir/weights directory.
docker run --runtime=nvidia -it -v <path to weight directory>:/workdir/weights ntech_i
Place the relevant model weights in the weights directory.
from ensemble import Ensemble
submit = Ensemble(detector_weights_path, video_sequence_weights_path,
first_video_face_weights_path, second_video_face_weights_path)
prediction = submit.inference(video_path)