torchvision-enhance

torchvision-enhance is used to enhance the offical PyTorch vision library torchvision. Here is the enhanced parts:

• support multi-channel(> 4 channels, e.g. 8 channels) images
• support 16-bit TIF file
• more easier to semantic segmentation transform

Support transforms

• RandomFlip
• RandomVFlip
• RandomHFlip
• RandomRotate
• RandomShift
• RandomCrop
• CenterCrop
• Resize
• Pad
• GaussianBlur
• PieceTransform
• Lambda
• ToTensor
• Normalize

Install

pip install torchvision-enhance

or install from the source

git clone 
pip install -r requirements.txt
python setup.py install

Dependencies

• numpy
• scipy
• Pillow
• PyTorch
• opencv
• scikit-image

Usage

For more useage, check out the example-classification.py and example-segmentation.py

from torchvision_x.datasets import image_loader
from torchvision_x.transforms import transforms_seg,functional

transform = transforms_seg.SegCompose([
#     transforms_seg.SegFlip(),
    transforms_seg.SegVFlip(), 
    # transforms_seg.SegHFlip(),
    # transforms_seg.SegRandomFlip(),
    # transforms_seg.SegRandomRotate(90),
    # transforms_seg.SegRandomShift(40),
    # transforms_seg.SegRandomCrop((256,256)),
    # transforms_seg.SegCenterCrop(224),
    # transforms_seg.SegResize(224),
    # transforms_seg.SegPad(20),
    # transforms_seg.SegNoise(dtype='uint16', var=0.001),  #TODO
    # transforms_seg.SegGaussianBlur(sigma=2, dtype='uint8', multichannel=False),
    # transforms_seg.SegPieceTransform(),
#     transforms_seg.SegLambda(lambda x: functional.to_tensor(x))
    transforms_seg.SegToTensor(),
    transforms_seg.SegNormalize((0.5,0.5,0.5),(0.5,0.5,0.5)),
    ])

trainset = image_loader.SemanticSegmentationLoader(
    rootdir='sample-data/', lstpath='sample-data/segmentation_jpg.lst',
    filetype='jpg', transform=transform,
    )
trainloader = DataLoader(dataset=trainset,batch_size=batch_size,shuffle=False)

for step, (inputs, targets) in enumerate(trainloader):
    print('batch: {} ........'.format(step))
    print(type(inputs), inputs.shape)
    print(type(targets), targets.shape)

TODO

• Noise