My solution for the Deepfake Detection Challenge
Assuming the user has a standard Anaconda3 environment
-
Install NVIDIA DALI. Instructions can be found here.
-
Create a new folder called
vendorsor something similar. It will be used for 3rd party code which lacks proper install mechanisms -
Install Pytorch_Retinaface:
cd vendors
git clone https://github.com/biubug6/Pytorch_Retinaface.gitthen go to src/prepare_data.py and update this line with path to Pytorch_Retinaface in your system:
sys.path.insert(0, '/home/dmitry/projects/dfdc/vendors/Pytorch_Retinaface')-
Download pretrained weights from Google Drive.
-
Install other non-standart pip packages:
pip install h5py numba pytorch-ignite hydra-core