Hey, I’ve been trying to use the DataParallel class for training on several local GPUs. The issue I have is that my input data doesn’t fit the type of data that DataParallel, and in particular its scatter function, expects.
The scatter function in DataParallel recursively looks for Tensor types in the input data and distributes it between your GPUs along the axis you tell it to. This is all well and fine but it doesn’t consider the use case of people using lists of Tensors as minibatches. Is this a use case that will be supported, or is there a way to customize the scatter function myself?