search

vmazashvili / Neural-Networks

Masktune Project reimplementation
0 stars 0 forks source link

new masking techniques #6

Closed vmazashvili closed 2 weeks ago

vmazashvili commented 2 weeks ago

Ideas:

  1. Top-k Saliency Masking
  2. Channel-wise Masking
  3. Guided Backpropagation Masking
  4. Random Erasing
  5. CutOut
vmazashvili commented 2 weeks ago

modify the code in a way that I can generate the datasets and save them in masked_train_dataloaders with the masking function and parameter identifier. for example, in fine tuning loop, ill be able to save each finetuned model checkpoint for each dataset torch.save(model.state_dict(), f'/content/data/MNIST/CNN_checkpointfinetuned{method}_{parameter}.pth')

vmazashvili commented 2 weeks ago

added: Top-k Saliency Masking Channel-wise Masking