[Bug] mmpretrain.datasets.transforms.GaussianBlur doesn't work

[liyunlong10]

Branch

main branch (mmpretrain version)

Describe the bug

When I use the GaussianBlur to filter the input images, the outputs are always the same as the inputs.

Here is a snippet to reproduce the problem.

import numpy as np
from mmpretrain.datasets.transforms import GaussianBlur

# Create a simple image
img_ori = np.zeros((128, 128), dtype=np.uint8)
img_ori[32:96, 32:96] = 255
results = {'img': img_ori.copy()}

# Run the GaussianBlur
blur = GaussianBlur(radius=10, prob=1.0)
results = blur.transform(results)
img_blur = results['img']

# `img_blur` is expected to be different from `img_ori`
# but actually `img_blur` is exactly equal to `img_ori`
np.all(img_blur == img_ori) # False expected, but got True!

Environment

{'sys.platform': 'linux',
 'Python': '3.10.9 (main, Mar  1 2023, 18:23:06) [GCC 11.2.0]',
 'CUDA available': True,
 'numpy_random_seed': 2147483648,
 'GPU 0,1,2,3': 'Tesla K80',
 'CUDA_HOME': '/root/miniconda3/envs/py10env',
 'NVCC': 'Cuda compilation tools, release 11.7, V11.7.99',
 'GCC': 'gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0',
 'PyTorch': '1.13.1',
 'TorchVision': '0.14.1',
 'OpenCV': '4.7.0',
 'MMEngine': '0.8.1',
 'MMCV': '2.0.1',
 'MMPreTrain': '1.0.0+'}

Other information

I think the image is blurred, but the result is not properly saved.

The GaussianBlur class is in "mmpretrain/mmpretrain/datasets/transforms/auto_augment.py".
In the GaussianBlur.transform method, the blurred image returned by the following code is not properly saved,

pil_img = Image.fromarray(img)
pil_img.filter(ImageFilter.GaussianBlur(radius=radius))  # This is not an in-place operation.
results['img'] = np.array(pil_img, dtype=img.dtype)

It should be

pil_img = Image.fromarray(img)
pil_img = pil_img.filter(ImageFilter.GaussianBlur(radius=radius))  # The result should be saved
results['img'] = np.array(pil_img, dtype=img.dtype)

[Ezra-Yu]

Yes, it is!

Thanks for reporting this BUG， Could you create a PR and fix this BUG?