training data also normalized?

[hli2020]

Hi G,
I am so new to Torch, just a quick concern about fetching training data. The training data is supposed to be normalized, too. However, I see no such operation in the dataTrain:getBatch() call. Specifically, the code here does not pass input value back to batch. Can you point out where I misunderstood? Thanks!

[hughperkins]

If I'm reading correctly:

• this line https://github.com/gcr/torch-residual-networks/blob/master/data/cifar-dataset.lua#L58 creates a reference variable input, that points to one image, in the batch.inputs tensor
• this line https://github.com/gcr/torch-residual-networks/blob/master/data/cifar-dataset.lua#L74 , ie the one you point to:
  • flips the input, ie the image pointed to by input, which is in the batch.inputs tensor
  • copies the result of flipping, back into the batch.inputs tensor, via the input reference variable

(Edit: actually, input is not a refernce variable, it's actually a brand-new torch tensor, but that tensor has the exact same underlying storage as the origianl tensor, and any changes to the data in the input tensor write through to the exact same storage in the original tensor. this is probably the new information you are looking for?)

[gcr]

Yes, that's right. Perhaps this should have been made more clear with a comment or something: input refers to the same memory as batch.inputs, so mutating the values at input will also propagate to batch.input