IE643CourseProject

Face Manipulation Detection

The following reference was used:
N. Bonettini, E. D. Cannas, S. Mandelli, L. Bondi, P. Bestagini and S. Tubaro, "Video Face Manipulation Detection Through Ensemble of CNNs," 2020 25th International Conference on Pattern Recognition (ICPR), 2021, pp. 5012-5019, doi: 10.1109/ICPR48806.2021.9412711.
The github repo: https://github.com/polimi-ispl/icpr2020dfdc#datasets

Suitable changes have been made to run it on NVIDIA GEFORCE GTX 1660 TI GPU, intel core i7 processor.

Install pytorch from official website in a conda environment
Update the environment using environment.yml file commenting pytorch wherever given
Original code uses python 3.6.9 but to utilize GPU in training python version 3.7.x has been used in this code

Several depracated functions have been replaced with the new versions

Tweak the learning rate and max_iter in .sh files for a faster training of the model
If Cuda goes out of memory, try lowering down the batch size of training dataset

The results obtained by testing the trained model [stored in weight directory] is stored in results directory.
The bestval.pth in weights directory is the file with the minimum loss encountered during training procedure

Pre trained models have also been preented as usable notebooks in notebook directory

Nikhil Tiwari
200010050
In guidance of Prof P Balamurugan