Basic: Deep learning with PyTorch

If you followed through the guide, you may have noticed that bioimageloader has many classes whose interfaces are similar to those of PyTorch, a popular python library for deep learning. It is true that bioimageloader followed the design of PyTorch’s data module, and especially so for bioimageloader.ConcatDataset and bioimageloader.BatchDataloader. The benefit is that all bioimageloader datasets are compatible with PyTorch.

For example, you can replace bioimageloader.BatchDataloader with its equivalent DataLoader from PyTorch, while keeping the same transformation API from albumentations. In fact, albumentations is a part of the PyTorch ecosystem

Note that you may want to convert image arrays to Tensor which PyTorch can handle, using ToTensorV2 from albumentations library. Find a full example guide here.

import albumentations as A
from albumentations.pytorch import ToTensorV2  # PyTorch helper
from bioimageloader.collections import DSB2018
from torch.utils.data import DataLoader  # PyTorch

transforms_totensor = A.Compose([
    A.RandomCrop(width=256, height=256),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
    ToTensorV2(),  # convert numpy ndarray to torch Tensor
])
# construct dset with transforms
dsb2018 = DSB2018('./data/DSB2018', transforms=transforms_totensor)
batch_loader = Dataloader(dsb2018,
                          batch_size=16,
                          drop_last=True,
                          num_workers=8)
# iterate transformed images
data: dict[str, numpy.ndarray]
for data in dsb2018:
    image = data['image']  # torch.Tensor, (b, c, h, w)
    mask = data['mask']  # torch.Tensor, (b, h, w)
    # these assertions will not throw AssertionError
    assert image.shape[0] == 16 and mask.shape[0] == 16
    assert image.shape[1] == 3  # notice Tensor convention shape (b, c, h, w)
    assert image.shape[2] == 256 and image.shape[3] == 256
    assert mask.shape[1] == 256 and mask.shape[2] == 256
    do_something(image, mask)