figure_masking_threshold.py

import dataclasses
from itertools import product
import pathlib

import matplotlib.pyplot as plt

from benchmarking.distance_benchmark import run_image_vs_image_experiment, ImageVsImageCase, log_git_versions, plot_saliency_map_on_image, run_image_captioning_experiment, ImageCaptioningCase
from benchmarking.Config import original_config_options, Config


imagenet_case = ImageVsImageCase(name='bee_vs_fly',
                                 input_image_file_name='bee.jpg',
                                 reference_image_file_name='fly.jpg')

image_size = 224
half_image_size = image_size / 2

fancy_figure_kwargs = {
    # much fun with DPI, column width and font size (and font type of course!)
    # ... once we know these things
    'alpha': 0.7
}

base_output_folder = pathlib.Path('paper_figures')


mask_threshold_configs = [dataclasses.replace(original_config_options,
                                              experiment_name=f'mask_selection_thresholds_neg{neg_min}-{neg_max}_pos{pos_min}-{pos_max}',
                                              mask_selection_range_max=pos_max,
                                              mask_selection_range_min=pos_min,
                                              mask_selection_negative_range_max=neg_max,
                                              mask_selection_negative_range_min=neg_min,
                                              ) for pos_min, pos_max, neg_min, neg_max in
                          [
                              (0.0, 0.05, 0.95, 1.0),
                              (0.0, 0.1, 0.9, 1.0),
                              (0.0, 0.15, 0.85, 1.0),
                              (0.0, 0.2, 0.8, 1.0),
                              (0.0, 0.4, 0.6, 1.0),
                          ]]


def make_figure():
    fig, ax = plt.subplots(1, 5, figsize=(12, 2), layout="constrained")

    for ix, config in enumerate(mask_threshold_configs):
        output_folder = base_output_folder / f'{config.experiment_name}'
        output_folder.mkdir(exist_ok=True, parents=True)
        config.to_yaml_file(output_folder / 'config.yml')

        log_git_versions(output_folder)

        case_folder = output_folder / f'image_vs_image_{imagenet_case.name}'
        case_folder.mkdir(exist_ok=True, parents=True)
        saliency, central_value, input_image = run_image_vs_image_experiment(imagenet_case, config, case_folder, analyse=False)

        ax_ix = ax.flatten()[ix]

        plot_saliency_map_on_image(input_image, saliency[0], ax=ax_ix,
                                   title="", add_value_limits_to_title=False,
                                   vmin=saliency[0].min(), vmax=saliency[0].max(),
                                   central_value=central_value, **fancy_figure_kwargs)
        
        prepend = ""
        selected = round((config.mask_selection_negative_range_max - config.mask_selection_negative_range_min) * 100)
        if ix == 0:
            prepend = "masks selected (each side): "
        ax_ix.text(half_image_size, image_size + 20, f'{prepend}{selected}%',
                   horizontalalignment='center', verticalalignment='center')

    fig.savefig(base_output_folder / 'masking_threshold.pdf')



if __name__ == '__main__':
    make_figure()