PyTorch’s default dataloader tends to get annoying, especially when we deal with custom datasets/conditional dataset loading. For a project we were working on we had to load a number of large datasets that weren’t structured the way the ImageFolder DataLoader expects, so we modified it to allow the user to specify whatever structure they want. It’s been a cool experiment so far, but we’re contemplating taking it on as a long-term open source project, and I was wondering if this problem is as common as we feel it would be? Would love to hear what people think, the code is at https://github.com/BinItAI/BetterLoader, if you want to check it out
We're working on a tool to make loading image datasets in PyTorch less difficult | Heykuki News