Transformations

Code for this section：

core/data/dataloader.py
core/data/collates/contrib/__init__.py
core/data/collates/collate_functions.py

In LibFewShot，we use a base transforms to compare some methods fairly. The base transforms could be divided into three sub-transforms:

Resize&Crop + ExtraTransforms + ToTensor&Norm

There are some differences in Resize&Crop for different dataset and config file (key augment):

1. in the training phase, config.augment is True：

   from torchvision import transforms
   transforms.RandomResizedCrop((config.image_size, config.image_size))
   

2. in other phases：

   from torchvision import transforms
   transforms.Resize((96, 96))  # or 256 when config.image_size = 224
   transforms.CenterCrop((84, 84)) # or 224 when config.image_size = 224
   

Besides, you may notice that ToTensor & Norm always uses the same sets of mean and variance, then you can reset mean and variance for different datasets.

MEAN = [120.39586422 / 255.0, 115.59361427 / 255.0, 104.54012653 / 255.0]
STD = [70.68188272 / 255.0, 68.27635443 / 255.0, 72.54505529 / 255.0]