DeepFakeDetction

Hi CS231N TAs! Thank you for taking the time to evaluate our project. Please follow the instructions below to replicate the results from our final report and poster presentation.

We've excluded our trained model weights to keep this repository light. If you'd like them, please email me at colins26@stanford.edu, and I'd be happy to send them over.

Setup and Execution

• Make sure you have conda installed
• Create the environment with conda env create -f environment.yml
• Activate the environment with conda activate deepfakesonly
• Run python download.py to download data (you may need to enter your Kaggle username and API key)
  • To get an API key, go to the 'Account' tab of your user profile and select 'Create New Token'. This will trigger the download of kaggle.json, a file containing your API credentials.
• Run python train.py to train all models
  • If on a slurm cluster, you can run sbatch script.slurm to train models in parallel
• Run python analyze.py to evaluate all models and reproduce all plots

Tensorboard (optional)

• To view intermediate training results, run tensorboard --logdir=logs --port=6006 and open up https://localhost:6006 in a browser
• If on a gcloud compute cluster, you can run gcloud compute ssh [INSTANCE_NAME] -- -NfL 6006:localhost:6006 to forward the tensorboard output to your local machine