opencv_torchvision_transform

1. This is an opencv based rewriting of the "transforms" in torchvision package.

2. All functions depend on only cv2 and pytorch (PIL-free). As the article says, cv2 is three times faster than PIL.

3. Most functions in transforms are reimplemented, except that:

   1. ToPILImage (opencv we used :)), Scale and RandomSizedCrop which are deprecated in the original version are ignored.

   2. The affine transform in the original one only has 5 degrees of freedom, I implement an Affine transform with 6 degress of freedom called RandomAffine6 (can be found in cvtransforms.py). The original method RandomAffine is still retained and reimplemented with opencv.

   3. My rotate function is clockwise, however the original one is anticlockwise.

   4. Adding some new methods which can be found in Support (the bolded ones).

   5. All the outputs of the opencv version are almost the same as the original one's (test in cvfunctional.py).

Support:

• Compose, ToTensor, ToCVImage, Normalize

• Resize, CenterCrop, Pad

• Lambda (doesn't work well in multiprocess in Windows)

• RandomApply, RandomOrder, RandomChoice, RandomCrop,

• RandomHorizontalFlip, RandomVerticalFlip, RandomResizedCrop,

• FiveCrop, TenCrop, LinearTransformation, ColorJitter,

• RandomRotation, RandomAffine, *RandomAffine6, *RandomPerspective

• *RandomGaussianNoise, *RandomPoissonNoise, *RandomSPNoise

• Grayscale, RandomGrayscale

How to use:

1. git clone https://github.com/YU-Zhiyang/opencv_torchvision_transforms.git .

2. Add cvtorchvision to your python path.

3. Add from cvtorchvision import cvtransforms in your python file.

4. You can use all functions as the original version, for example:

   transform = cvtransforms.Compose([
    
            cvtransforms.RandomAffine(degrees=10, translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 0),
    
            cvtransforms.Resize(size=(350, 350), interpolation='BILINEAR'),
    
            cvtransforms.ToTensor(),
    
            cvtransforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
            ])
   

more details can be found in the examples of official tutorials.

Good News:

You can install this package via pip install opencv-torchvision-transforms-yuzhiyang (Old version only)

Attention:

The multiprocessing used in dataloader of pytorch is not friendly with lambda function in Windows as lambda function can't be pickled (https://docs.python.org/3/library/pickle.html#what-can-be-pickled-and-unpickled).

So the Lambda in cvtransforms.py may not work properly in Windows.

Requirements

python >=3.5.2

numpy >=1.10 ('@' operator may not be overloaded before this version)

pytorch>=0.4.1

torchvision>=0.2.1

opencv-contrib-python-3.4.2 (test with this version, but any version of opencv3 is ok, I think)

Postscript

Welcome to point out and help fixing bugs!

Thanks HongChu who helps a lot.