FOSMix

Frequency-based Optimal Style Mix

Step 0

Install packages

$ pip install pipenv
$ virtualenv -p ~/.pyenv/versions/3.10.4/bin/python ~/venvs/fosmix
$ source ~/venvs/fosmix/bin/activate
$ (fosmix) pip install -r requirements.txt

Step 1

Modify the installed packages

$ (fosmix) chmod 777 modify_package_contents.sh
$ (fosmix) ./modify_package_contents.sh

Step 2

Train the model

$ (fosmix) ./train.sh

Argument Description

Hyper parameters with * is required.

1. --dataset (str*) : Select the dataset.

   • OEM (OpenEarthMap)
   • FLAIR

2. --n_epochs (int*): Number of training epochs.

   • 150 (for OEM dataset)
   • 50 (for FLAIR dataset)

3. --ver (int*): Version.

4. --final (bool): Use the final model parameters for testing.

   • 0 (Use the parameters that gave the best results on the validation data for testing)
   • 1 (Use the last parameters)

5. --randomize (bool*): Randomize images ot not, i.e., whether to use the baseline or the proposal.

   • 0 (baseline)
   • 1 (use the proposal)

6. --optimize (bool*): Optimize the mask or not. When optimize is 1, randomize is always 1.

   • 0 (baseline or FULL MIX)
   • 1 (OPTIMAL MIX)

7. --aug_color (float): Probability of color change in augmentation.

8. --MFI (int*): Mask From Image or not, whether to generate the OPTIMAL MASK from the image.

   • 0 (Learn and use one mask for all images)
   • 1 (Generate from the image)

9. --fullmask (int*): Use FULL MIX or not. You can Also use the optimize together with this option. If fullmask is 1, randomize must always be set to 1 as well.

   • 0 (Do not use FULL MIX)
   • 1 (Use FULL MIX)