salad.datasets package

Datasets for Domain Adaptation Experiments.

This package contains datasets and tools for handling datasets. Similar as in torchvision.datasets, data is accessed through subclasses of torch.utils.data.DataLoader and torch.utils.data.Dataset.

As one very established Domain Adaptation benchmarks, the digits package focusses on the small digit benchmark consisting of

• MNIST
• USPS
• SVHN
• SYNTH

Principally two main methods for loading data are currently implemented. In general, multiple datasets are loaded.

In cat mode, the dataset returns values of the form

>>> for x,y,d in data_loader:
>>>     print(x.size(), y.size(), d.size())

In stack mode, the dataset returns tuples (of possible different sizes):

>>> for (xs,ys), (xt, yt) in data_loader:
>>>     pass

Subpackages

• salad.datasets.da package
  • Submodules
  • salad.datasets.da.base module
  • salad.datasets.da.digits module
  • salad.datasets.da.toy module
• salad.datasets.digits package
  • Submodules
  • salad.datasets.digits.base module
  • salad.datasets.digits.mnist module
  • salad.datasets.digits.openset module
  • salad.datasets.digits.synth module
  • salad.datasets.digits.usps module
• salad.datasets.transforms package
  • Submodules
  • salad.datasets.transforms.digits module
  • salad.datasets.transforms.ensembling module
  • salad.datasets.transforms.noise module
  • salad.datasets.transforms.noisy module
• salad.datasets.visda package
  • Submodules
  • salad.datasets.visda.detection module
  • salad.datasets.visda.openset module
  • salad.datasets.visda.utils module
  • salad.datasets.visda.visda module