Skip to content
This repository has been archived by the owner on Oct 12, 2021. It is now read-only.
/ DeepLearning Public archive

Deeplearning playground for IMT Mines Ales deep learning course exercise

Notifications You must be signed in to change notification settings

BastLast/DeepLearning

Repository files navigation

Deep Learning project

Missions

You must find a way to restore/decrypt images!

Two challenges are given.

  1. Image restoration
  • Dataset 1A: images only need to be restored, examples of damaged and restored version of images are provided.
  • Dataset 1B: images have to be restored using advanced techniques, examples of damaged and restored version of images are also provided.
  1. Image decryption
  • Dataset 2: images are clearly encrypted... but some of them have been cracked!!! Examples of encrypted images and corresponding original images are provided.

Both train and test sets are provided.

Rules

  • You must use Deep Learning techniques.
  • Results must be reproductible, use torch.manual_seed(1234). Training must also be reproductible.
  • Team size <= 4
  • The following evaluation metric will be used.
def eval_metric(img, pred):
     return torch.abs(img - pred).sum() 
  • Do not use the test set during training... The test set cannot be used to train or select you model. test set + eval_metric can be used to compare your results with other teams.
  • For each challenge, a bonus will be given to the best team (the one maximizing eval_metric on the test set). Best team bonus will be +2/#number of team members (challenge 1, average between Dataset 1A, and 1B will be made).

You should submit:

  • Trained models.
  • A report detailing the methodology, tested architectures, results, illustration of predictions on the test set, as well as discussions. You must report the eval_metric results for the test sets (max 20 pages).

About

Deeplearning playground for IMT Mines Ales deep learning course exercise

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published