Hi
Does someone know if it is possible to load the MSCOCO 2017 dataset with the FederatedDataset class and partition it correctly? I have struggled to do so, since pytorch expect another format than the coco format. Therefore i was wondering if it possible to load a custom pytorch datasetclass created myself in the FederatedDataset, instead of using the string path of it from huggingface?
Hello @johanrubak, I haven’t experienced loading MSCOCO from HuggingFace before. But you can create a FederatedDataset from your own local images or in-memory data. I think there’s potentially one conversion step to convert PyTorch Dataset to HuggingFace’s Dataset. After that, you can create the FederatedDataset from the Dataset.