Corrupting Images with TorchUncertainty to Benchmark Robustness

This tutorial shows the impact of the different corruption transforms available in the TorchUncertainty library. These corruption transforms were first proposed in the paper Benchmarking Neural Network Robustness to Common Corruptions and Perturbations by Dan Hendrycks and Thomas Dietterich.

For this tutorial, we will only load the corruption transforms available in torch_uncertainty.transforms.corruption. We also need to load utilities from torchvision and matplotlib.

from torchvision.datasets import CIFAR10
from torchvision.transforms import Compose, ToTensor, Resize

import matplotlib.pyplot as plt

ds = CIFAR10("./data", train=False, download=True)


def get_images(main_corruption, index: int = 0):
    """Create an image showing the 6 levels of corruption of a given transform."""
    images = []
    for severity in range(6):
        ds_transforms = Compose(
            [ToTensor(), main_corruption(severity), Resize(256, antialias=True)]
        )
        ds = CIFAR10("./data", train=False, download=False, transform=ds_transforms)
        images.append(ds[index][0].permute(1, 2, 0).numpy())
    return images


def show_images(transforms):
    """Show the effect of all given transforms."""
    num_corruptions = len(transforms)
    _, ax = plt.subplots(num_corruptions, 6, figsize=(10, int(1.5 * num_corruptions)))
    for i, transform in enumerate(transforms):
        images = get_images(transform, index=i)
        ax[i][0].text(
            -0.1,
            0.5,
            transform.__name__,
            transform=ax[i][0].transAxes,
            rotation="vertical",
            horizontalalignment="right",
            verticalalignment="center",
            fontsize=12,
        )
        for j in range(6):
            ax[i][j].imshow(images[j])
            if i == 0 and j == 0:
                ax[i][j].set_title("Original")
            elif i == 0:
                ax[i][j].set_title(f"Severity {j}")
            ax[i][j].axis("off")
    plt.show()

Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ./data/cifar-10-python.tar.gz

  0%|          | 0/170498071 [00:00<?, ?it/s]
  0%|          | 32768/170498071 [00:00<11:49, 240269.91it/s]
  0%|          | 229376/170498071 [00:00<03:00, 941277.73it/s]
  1%|          | 884736/170498071 [00:00<01:02, 2696865.10it/s]
  2%|▏         | 3538944/170498071 [00:00<00:16, 9862902.72it/s]
  4%|▍         | 7536640/170498071 [00:00<00:08, 19017600.61it/s]
  7%|▋         | 12681216/170498071 [00:00<00:05, 26559406.37it/s]
 11%|█         | 18087936/170498071 [00:00<00:04, 34365527.81it/s]
 13%|█▎        | 22020096/170498071 [00:00<00:04, 35329731.99it/s]
 16%|█▌        | 26443776/170498071 [00:01<00:03, 37894524.52it/s]
 18%|█▊        | 31391744/170498071 [00:01<00:03, 39832059.53it/s]
 21%|██        | 35454976/170498071 [00:01<00:03, 40006637.94it/s]
 24%|██▍       | 40632320/170498071 [00:01<00:02, 43391956.41it/s]
 26%|██▋       | 45023232/170498071 [00:01<00:03, 41434029.47it/s]
 29%|██▉       | 50069504/170498071 [00:01<00:02, 43690261.21it/s]
 32%|███▏      | 54493184/170498071 [00:01<00:02, 42085256.10it/s]
 35%|███▍      | 59539456/170498071 [00:01<00:02, 44393856.86it/s]
 38%|███▊      | 64028672/170498071 [00:01<00:02, 42472037.59it/s]
 41%|████      | 69107712/170498071 [00:02<00:02, 44668573.40it/s]
 43%|████▎     | 73629696/170498071 [00:02<00:02, 42706653.94it/s]
 46%|████▌     | 78643200/170498071 [00:02<00:02, 44385731.98it/s]
 49%|████▉     | 83132416/170498071 [00:02<00:02, 42839839.10it/s]
 52%|█████▏    | 88047616/170498071 [00:02<00:01, 44588127.70it/s]
 54%|█████▍    | 92569600/170498071 [00:02<00:01, 42715475.08it/s]
 57%|█████▋    | 97681408/170498071 [00:02<00:01, 44754933.70it/s]
 60%|█████▉    | 102203392/170498071 [00:02<00:01, 42944401.48it/s]
 63%|██████▎   | 107216896/170498071 [00:02<00:01, 44803003.80it/s]
 66%|██████▌   | 111738880/170498071 [00:03<00:01, 42825947.44it/s]
 68%|██████▊   | 116588544/170498071 [00:03<00:01, 44279430.87it/s]
 71%|███████   | 121077760/170498071 [00:03<00:01, 42780488.57it/s]
 74%|███████▍  | 126025728/170498071 [00:03<00:01, 44169653.44it/s]
 77%|███████▋  | 130482176/170498071 [00:03<00:00, 42763713.55it/s]
 79%|███████▉  | 135397376/170498071 [00:03<00:00, 44392883.67it/s]
 82%|████████▏ | 139886592/170498071 [00:03<00:00, 42716394.35it/s]
 85%|████████▌ | 144965632/170498071 [00:03<00:00, 44961557.83it/s]
 88%|████████▊ | 149520384/170498071 [00:03<00:00, 42722400.83it/s]
 91%|█████████ | 154697728/170498071 [00:03<00:00, 44637788.65it/s]
 93%|█████████▎| 159219712/170498071 [00:04<00:00, 43045896.42it/s]
 96%|█████████▌| 164003840/170498071 [00:04<00:00, 44304719.74it/s]
 99%|█████████▉| 168493056/170498071 [00:04<00:00, 42312578.95it/s]
100%|██████████| 170498071/170498071 [00:04<00:00, 39171789.70it/s]
Extracting ./data/cifar-10-python.tar.gz to ./data

1. Noise Corruptions

from torch_uncertainty.transforms.corruption import (
    GaussianNoise,
    ShotNoise,
    ImpulseNoise,
    SpeckleNoise,
)

show_images(
    [
        GaussianNoise,
        ShotNoise,
        ImpulseNoise,
        SpeckleNoise,
    ]
)

2. Blur Corruptions

from torch_uncertainty.transforms.corruption import (
    GaussianBlur,
    GlassBlur,
    DefocusBlur,
)

show_images(
    [
        GaussianBlur,
        GlassBlur,
        DefocusBlur,
    ]
)

3. Other Corruptions

from torch_uncertainty.transforms.corruption import (
    JPEGCompression,
    Pixelate,
    Frost,
)

show_images(
    [
        JPEGCompression,
        Pixelate,
        Frost,
    ]
)

Downloading https://zenodo.org/records/10438904/files/frost.zip to data/frost.zip

  0%|          | 0/1780178 [00:00<?, ?it/s]
  2%|▏         | 32768/1780178 [00:00<00:07, 218593.56it/s]
  4%|▎         | 65536/1780178 [00:00<00:07, 220783.51it/s]
  7%|▋         | 131072/1780178 [00:00<00:05, 322492.91it/s]
 17%|█▋        | 294912/1780178 [00:00<00:02, 631278.85it/s]
 33%|███▎      | 589824/1780178 [00:00<00:01, 1120540.67it/s]
 66%|██████▋   | 1179648/1780178 [00:00<00:00, 2092604.30it/s]
100%|██████████| 1780178/1780178 [00:00<00:00, 1994209.55it/s]
Extracting data/frost.zip to data

Reference

• Benchmarking Neural Network Robustness to Common Corruptions and Perturbations, Dan Hendrycks and Thomas Dietterich. ICLR 2019.