Up to now I’ve figure out two approaches to use a custom dataset(with the standard DataLodar).
The first is to use ImageFolder, somehow similar to creating a filelist and get access to data. Transformations such as randomcrop and scaling is convenient with that scheme,but I guess it is I/O-consuming.
dataset = trochvision.datasets.ImageFolder(root=‘main_dir’,transformer=torchvision.transform.Compose([…])
So the second scheme is to read in a bunch of data into the memory at a time into tensors:
dataset = torch.utils.data.TensorDataset(data_tensor,label_tensor)
Anything like the transformer in the first approach for TensorDataset?