The original deepfake dataset consists of 50 folders with ~3000 videos per folder with a metadata.json file with labels for every video. Classes are imbalanced: 80% of videos are fake and 20% are real. You can find more information about the dataset on the web page of the challenge.
This dataset was pre processed in the nbs/Colab_ds_preproc.ipynb notebook in order to create a balanced dataset with proper validation.
The project is structured as following:
.
├── datasets
| └── dataset selector
| └── video_dataset.py # creating a PyTorch video dataset
├── models
| └── architecture selector
| └── resnet_rnn.py # CNN with a following RNN
├── nbs
| └── 00_video_dataset.ipynb # notebook used to create the dataset script
| └── 01_model.ipynb # notebook used to create the resnet_rnn script
| └── 02_train.ipynb
| └── Colab_ds_preproc.ipynb # preprocessing of the original dataset in Google Colab
| └── notebook2script.py
| └── train_fastai.py # fastai model was trained using this script
├── args.py # parsing all command line arguments for experiments
├── trainer.py # main file from the project serving for calling all necessary functions for training and testingExperiments can be launched by calling train.py and a set of input arguments to customize the experiments. You can find the list of available arguments in args.py and some default values. Note that not all parameters are mandatory for launching and most of them will be assigned their default value if the user does not modify them.
Here is a typical launch command and some comments:
python train.py --batch_size 16 --crop_size 224 --csv_file videos_df --dataset video --dropout 0.3 --embed_dim 512 --epochs 50 --log_interval 10 --lr 0.001 --model_name resnet152 --num_classes 2 --pretrained True --workers 4
For each experiment weights of all the networks are saved to the folder checkpoints after each epoch as well as all the train and valiation losses and accuracy values.