Please install PyTorch as appropriate for your system. This codebase has been developed with python 3.8.12, PyTorch 1.11.0, CUDA 11.3 and torchvision 0.12.0 with the use of an anaconda environment.
To create an appropriate conda environment (after you have successfully installed conda), run the following command:
conda create --name fewture --file requirements.txt
Activate your environment via
conda activate fewture
To download the miniImageNet dataset, you can use the script download_miniimagenet.sh in the datasets folder.
The miniImageNet dataset (Vinyals et al., 2016; Ravi & Larochelle, 2017) consists of a specific 100 class subset of Imagenet (Russakovsky et al., 2015) with 600 images for each class. The data is split into 64 training, 16 validation and 20 test classes.
To download the tieredImageNet dataset, you can use the script download_tieredimagenet.sh in the datasets folder.
Similar to the previous dataset, the tieredImageNet (Ren et al., 2018) is a subset of classes selected form the bigger ImageNet dataset (Russakovsky et al., 2015), however with a substantially larger set of classes and different structure in mind. It comprises a selection of 34 super-classes with a total of 608 categories, totalling in 779,165 images that are split into 20,6 and 8 super-classes to achieve better separation between training, validation and testing, respectively.
To download the CIFAR-FS dataset (Bertinetto et al., 2018), you can use the script download_cifar_fs.sh in the datasets folder.
The CIFAR-FS dataset contains the 100 categories with 600 images per category from the CIFAR100 dataset (Krizhevsky et al., 2009) which are split into 64 training, 16 validation and 20 test classes.
To download the FC-100 dataset, you can use the script download_fc100.sh in the datasets folder.
The FC-100 dataset (Oreshkin et al., 2018) is also derived from CIFAR100 (Krizhevsky et al., 2009) but follows a splitting strategy similar to tieredImageNet to increase difficulty through higher separation, resulting in 60 training, 20 validation and 20 test classes.