Awesome

The Hitchhiker's Guide to Prior-Shift Adaptation

Suplementary material

This sample code for prior shift adaptation in supplementary_material directory is a supplementary material for the paper:

Tomas Sipka, Milan Sulc and Jiri Matas. The Hitchhiker's Guide to Prior-Shift Adaptation. WACV, 2022.

See supplementary_material/example.ipynb for the usage of the methods and their comparison.

Prior-shift package

The code in prior_shift directory can be used as python package, providing a convenient adapt_to_prior_shift function for adapting your predictions to prior shift. See the sample usage in sample.ipynb.

Installation

Requirements:

numpy
sklearn

Copy prior_shift directory into your project and import it.

Usage

You can adapt your predictions with default settings:

import prior_shift as ps
test_pred_adapted = ps.adapt_to_prior_shift(test_pred, val_pred, val_targets, train_pred)

or you can call the function with diffenrent arguments to change the algorithm or its hyperparameters:

adapt_to_prior_shift(test_pred, val_pred, val_targets, train_pred, train_targets, method, args)

Arguments:
  test_pred: np.array (n_test, num_classes), predictions on test set, which should be adapted
  val_pred: np.array (n_val, num_classes), predictions on validation set
  val_targets: np.array (n_val,), ground truth labels coresponding to the predictions on 
      validation dataset
  train_pred: np.array (n_train, num_classes), predictions on trainning set
  train_targets: np.array (n_train,), ground truth labels coresponding to the predictions 
      on trainning dataset
  method: str, the algorithm used for priors estimation, allowed values are "EM", "CM", "SCM", 
      "CM-L", "SCM-L", "CM-M", "SCM-M", as default SCM-L is used
  args: Dict[str, Any], parameters used for the algorithm (see below)

Method's parameters:

EM

• "termination_difference" (float), default=0.001
• "trainset_prior" ({"classif", "count"}), default="classif"

CM

• "trainset_prior" ({"classif", "count"}), default="classif"

SCM

• "trainset_prior" ({"classif", "count"}), default="classif"

CM-L

• "lr" (float), default=1e-4<br/>
• "max_iter" (integer), default=1000
• "trainset_prior" ({"classif", "count"}), default="classif"

SCM-L

• "lr" (float), default=1e-3<br/>
• "max_iter" (integer), default=1000
• "alpha" (float), default=3
• "trainset_prior" ({"classif", "count"}), default="classif"

CM-M

• "lr" (float), default=1e-3<br/>
• "max_iter" (integer), default=1000
• "trainset_prior" ({"classif", "count"}), default="classif"

SCM-M

• "lr" (float), default=1e-3<br/>
• "max_iter" (integer), default=1000
• "alpha" (float), default=3
• "trainset_prior" ({"classif", "count"}), default="classif"

Notes

• EM algorithm does not require validation set (val_pred and val_targets can be set to None).
• Validation set can have arbitrary distribution.
• Using training set instead of validation set can lead to decreased performance due to overfitting.
• If trainset_prior="classif" then train_targets can be set to None
• If trainset_prior="count" then train_pred can be set to None

Citation

If you use the code please cite the paper:

@InProceedings{Sipka_2022_WACV,
    author    = {\v{S}ipka, Tom\'a\v{s} and \v{S}ulc, Milan and Matas, Ji\v{r}{\'\i}}},
    title     = {The Hitchhiker's Guide to Prior-Shift Adaptation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2022},
    pages     = {1516-1524}
}