gan-aug-analysis

resources -

• weights of GANs:
  • soon To use these weights, please refer to each of the original implementations of PGAN, StyleGAN2, pix2pixHD, and SPADE.
• synthetic images used:
  • https://drive.google.com/drive/folders/10-rLZIuZ1ek2HZmN1jiQdT15x34Px6ea?usp=sharing

Train splits used:

Anonymization experiments:

Each different percentage considered is inside the directories, where the directory name represents the percentage used: ./journal/percentages/[0_0625, 0_125, 0_25, 0_5, 1_0]

• Inside each directory, we have a directory for each of the three cases considered:
  • In the root directory, the baseline, without synthetic.
  • real_sgan2, with ratio [1/x : 1/x] (doubling the training set).
  • real_14805_sgan2, with ratio [1/x : x - 1/x] (topping up the training set).

Augmentation:

• Experiments with different GANs:
  • ./splits/different_gans/[real-all, real-pgan, real-pix2pixhd, real-spade], ./splits/percentages/1_0/real_50p_sgan2>
• Experiments with different ratios of synthetics:
  • ./splits/percentages/1_0/[real_25p_sgan2, real_50p_sgan2, real_100p_sgan2]
• Experiments with different sampling methods:
  • ./splits/percentages/1_0/[real_50p_sgan2 (random), real_50p_sortedsgan2 (best), real_50p_reversesortedsgan2 (worst), real_50p_sgan2diverse (diverse]
  • ./splits/percentages/1_0/[real_100p_sgan2 (random), real_100p_sortedsgan2 (best), real_100p_reversesortedsgan2 (worst), real_100p_sgan2diverse (diverse]
• Experiments altering the malignant ratio:
  • ./splits/percentages/1_0/[real_50p_sgan2, real_50b75m_sgan, real_50b100m_sgan2]

Test sets csvs:

• derm7pt (clinical): .splits/testsets/derm7pt_clinical.csv
• derm7pt (dermato): .splits/testsets/derm7pt_dermato.csv
• isic19: .splits/testsets/isic2019_test.csv
• isic20: .splits/testsets/isic2020_subset_test.csv
• dermofit: .splits/testsets/dermofit_test.csv

Reproducing our work:

• We include scripts to execute all experiments: run_augmentation.sh and run_anonymization.sh
  • Modify it to include the correct path to the images of all the necessary datasets.
  • We used comet and sacred to organize the huge amount of runs/results. If you don't want to make use of such libraries, modify both train_comet_csv.py and test_comet_csv.py files to remove the dependencies.
  • In the other hand, if you plan to use comet, insert your own API_key and workspace_name in the train_comet_csv.py file (in the main function), and at the run_*.sh script.
  • Check if the images in the splits csv files have the intended path.