How to Ensure Reproducibility?

[kazuma-jp]

I'm testing reproducibility using the following code, but not all transformations produce consistent results.

import albumentations as A
import numpy as np
import random

def apply_transform(image, transform, seed):
    random.seed(seed)
    np.random.seed(seed)

    return transform(image=image)["image"]

check_transforms = [
    A.Blur(),
    A.HorizontalFlip(p=0.5),
    A.VerticalFlip(p=0.5),
]

image = 128 * np.eye(64, dtype=np.uint8)[..., None].repeat(3, axis=2)

for aug in check_transforms:
    equal = True

    for seed in range(100):
        img1 = apply_transform(image, aug, seed)
        img2 = apply_transform(image, aug, seed)

        equal = equal and (img1 == img2).all()

    print(f'{aug.__class__.__name__}:{"OK" if equal else "Mismatch"}')

Is this the expected behavior, or could this be an unexpected issue?

I'm using albumentations version 2.0.2.
Thank you for reading. I would appreciate any insights you can share!

[ternaus]

You should fix seed in Compose as a parameter.

transform = A.Compose([
    A.RandomCrop(height=256, width=256),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
], seed=42)

https://albumentations.ai/docs/faq/#how-to-have-reproducible-augmentations

[kazuma-jp]

Thank you!!