google-research
Model could use Mini-imagenet dataset?
mini-ImageNet suffers from several limitations, that we outline in the ICLR article, and has limited usefulness nowadays.
Moreover, the Meta-Dataset benchmark includes the full ImageNet with 1000 classes, of which the 64 classes of mini-ImageNet are a subset, but with a different (meta-)split. In particular, 12 training classes of ilsvrc_2012 (and 6 validation classes) are test classes of mini-Imagenet, and 11 training classes of mini-Imagenet (and 10 validation classes) are test classes of ilsvrc_2012, so both datasets cannot be part of the same benchmark.
mini-ImageNet can still be used as a sanity check for the implementation of algorithms, that is why there is some code to convert it to tfrecords in meta_dataset/dataset_conversion/dataset_to_records, and some Gin configuration files in meta_dataset/learn/gin.
