Hi, I noticed you fixed a bug of Resize, commit id: 2f6a07d. However, the latest version on pip doesn't include this change. Could you please release a new version? Thx.

Image and ImageEnhance is not imported in functional.py

In five_crop()

1. get h,w from img shape wrongly.
2. crop the image in a wrong way.
   It should be as belowed

def five_crop(img, size):
    if isinstance(size, numbers.Number):
        size = (int(size), int(size))
    else:
        assert len(size) == 2, "Please provide only two dimensions (h, w) for size."
    h,w = img.shape[0:2]
    crop_h, crop_w = size
    if crop_w > w or crop_h > h:
        raise ValueError("Requested crop size {} is bigger than input size {}".format(size,
                                                                                      (h, w)))
    tl = crop(img, 0, 0, crop_h, crop_w)
    tr = crop(img, 0, w-crop_w, crop_h, crop_w)
    bl = crop(img, h - crop_h,0, crop_h, crop_w)
    br = crop(img, h - crop_h, w- crop_w, crop_h, crop_w)
    center = center_crop(img, (crop_h, crop_w))
    return [tl, tr, bl, br, center]

Hi, I am curious that how you determine the number 74 in the line table = np.array([(i - 74) * contrast_factor + 74 in functional.py? Thank you.

first of all. thanks for work. it is what i want.
the opencv _transform snippet output

def custom_transform(path):
    img = cv2.cvtColor(cv2.imread(path), cv2.COLOR_BGR2RGB)
    img_resize = transforms.Resize(size=(224,224))(img)
    img_tensor = transforms.ToTensor()(img_resize)
    img_norm = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])(img_tensor)
    out = np.expand_dims(img_norm, 0)
    print(out[0][0][0][0])

the pytorch transform output

def torch_transform(path):
    transform_list = torch_transforms.Compose([
            torch_transforms.Resize(224),
            torch_transforms.ToTensor(),
            torch_transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225]),
        ])
    img = Image.open(path).convert('RGB')
    out = transform_list(img).unsqueeze_(0)
    print(out[0][0][0][0])

the result :
tensor(0.2453)
0.1939379

To follow the same call, and have an in-place replacement, we need pad = [pad[0], pad[2], pad[1], pad[3]]

I use your implementation of transforms to my datasets, but I find that Dataloader could not work, it is only support with PIL Image. STRANGE

Hi there, using cvtransforms.ColorJitter(brightness=(0, 0.1), contrast=(0, 0.1)) gives me error:
File "/home/ma/anaconda3/envs/ml/lib/python3.6/site-packages/opencv_transforms/transforms.py", line 710, in <lambda> transforms.append(Lambda(lambda img: F.adjust_brightness(img, brightness_factor)))' File "/home/ma/anaconda3/envs/ml/lib/python3.6/site-packages/opencv_transforms/functional.py", line 334, in adjust_brightness return cv2.LUT(img, table)[:,:,np.newaxis] cv2.error: OpenCV(4.2.0) /io/opencv/modules/core/src/lut.cpp:368: error: (-215:Assertion failed) (lutcn == cn || lutcn == 1) && _lut.total() == 256 && _lut.isContinuous() && (depth == CV_8U || depth == CV_8S) in function 'LUT'

Looking at these god awful opencv error messages, it might crash because the image is float32?

If the input image is grayscale， I think you should use len(image.shape) == 2 instead of image.shape[2] == 1

Great repo.
But it seems that adjust_hue and adjust_saturation functions are exactly the PIL implementations (PyTorch official implementations). I wonder are there OpenCV implementations for these two functions, since the adjust_hue is rather slow, and adjust_saturation is a little slow when compared with adjust_contrast and adjust_brightness functions.

In my testing cases, given one 224x224x3 input, adjust_contrast takes about 188 µs, adjust_saturation takes about 2.33 ms, and adjust_hue takes about 6.99 ms.

Good job

when I try to use the package:
from opencv_transforms import transforms
I get an error:

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
<ipython-input-71-d4430fa4d55e> in <module>
----> 1 from opencv_transforms import transforms

~/.local/lib/python3.8/site-packages/opencv_transforms/transforms.py in <module>
     16 # import opencv_functional as F
     17 import cv2
---> 18 from . import functional as F
     19 
     20 __all__ = [

~/.local/lib/python3.8/site-packages/opencv_transforms/functional.py in <module>
      2 import math
      3 import random
----> 4 from PIL import Image, ImageOps, ImageEnhance, PILLOW_VERSION
      5 try:
      6     import accimage

ImportError: cannot import name 'PILLOW_VERSION' from 'PIL' (/usr/lib64/python3.8/site-packages/PIL/__init__.py)

I've installed the package by:
pip install --user opencv_transforms
I have OpenCV package installed, as well as PIL.

According to Pillow documentation PILLOW_VERSION has been removed, and one is supposed to use __version__ now.
https://pillow.readthedocs.io/en/stable/releasenotes/7.0.0.html#pillow-version-constant

I am facing this error while reproducing the example code from https://pypi.org/project/opencv-transforms/.

python - 3.8.5
numpy - 1.19.1
opencv - 4.4.0

RESIZE = (128, 171)
MEAN = [0.43216, 0.394666, 0.37645]
STD = [0.22803, 0.22145, 0.216989]
CROP_SIZE = 112

transform = Compose(
                [
                    # Resize(RESIZE),
                    # CenterCrop(CROP_SIZE),
                    # ToTensor(),
                    Normalize(MEAN, STD)
                ]
            )

On trying the above code (even without commenting) on google collab I get the following error:

Does anyone know what the issue is or how to fix this? Attaching list of pip packages in colab for reference if needed packages.txt

Thanks

This is not an issue, just to let you know I've commited an update to my fork of your repo, that I combined with https://github.com/YU-Zhiyang/opencv_transforms_torchvision and additional augmentations using OpenCV from one of my repos.

I tried to keep the best from all sources, use only OpenCV or np functions and extend functionality where possible, like the resample flag on RandomRotation.

In case you are interested, here's the repo: https://github.com/victorca25/opencv_transforms should be able to easily pull from it, since I kept it as a fork.

Hi, I really appreciate your effort to share some valuable code.

My only question is whether using OpenCV is faster than Pillow-SIMD. I have read articles from the web that Pillow is surely slower than OpenCV, but referring to pages like https://python-pillow.org/pillow-perf/, there may be some cases when Pillow-SIMD is faster than OpenCV.

Have you tried using Pillow-SIMD instead of plain Pillow? If so, have you done any comparison between OpenCV and Pillow-SIMD on video datasets like Kinetics?

Thanks again.

opencv resize behave differently with pil resize. https://github.com/python-pillow/Pillow/issues/2718. Any idea to solve this?

Not sure if I'm doing something wrong, but this gives an error: "TypeError: 'int' object is not iterable"

transformations = transforms.Compose([
    transforms.Resize(255),
    # some other transforms
])