Closed MrDavidG closed 2 years ago
There is only a dataset_transform
argument for a MetaDataset
(the object containing the dataset). This is a transformation which is applied to a dataset; the major use-case in Torchmeta is to implement the split of data into train/test (support/query) sets. There is however no dataset_transform
argument in BatchMetaDataLoader
(the data-loader, which just iterates over the dataset).
Thanks for the library, it makes meta training much easier than before.
I wonder what's the meaning of 'dataset_transform' in
BatchMetaDataLoader
. According to the following toy example,dataset_transform
is like to specific the number of samples of tasks during training/testHowever, here is my example. I have set
test_shots
andshots
in apiminiimagenet
, so do I need to setdataset_transform
forBatchMetaDataLoader
?Also, does the argument
shuffle
have the same use in the both functions?