OWR-ImageClassification

Open World Recognition in Image Classification project for Machine Learning and Deep Learning course's assignment - PoliTO

Paper baseline implementation

• Finetuning
• Learning without forgetting (usage of distillation loss)
• iCarl (distillation + exemplars)

Ablation studies:

• Classifiers
  • Nearest Mean Exemplars (iCaRL standard) baseline
  • Fully connected layer
  • KNN
  • Cosine Layer + losses as detailed in

    Learning a Unified Classifier Incrementally via Rebalancing by Hue et all

• Losses (classification + distillation)
  • iCaRL (BCE + lfc) baseline
  • BCE + L1
  • BCE + L2
  • CE + L1
  • CE + L2

Open World setting:

• Naive rejection strategy
• Our variations

Proposed variation

Dataset usage - example

• Call the constructor Cifar100Dataset('split', transform)
• For each seed's iteration:
  • Call the function .define_splits(seed)
• For each split's iteration:
  • Call .change_subclasses(iteration) to select the list of 10 classes
  • Call .get_imgs_by_target() to retrieve all the images belonging to the 10 classes above, ready to fill the pytorch Subset.