I have the following setup for my data
data/ - 0000/ - 0000.png - 00001.png - ... - 0001/ - 0000.png - ... - ...
I’m trying to load 5 sequential frames so that one batch would return the following:
batch0: - 0000/0000.png, 0000/0001.png, ..., 0000/N.png - 0000/0001.png, 0000/0002.png, ..., 0000/N+1.png
I’m quite new to PyTorch and I feel a bit lost in different alternatives. I’ve looked at the following threads, but haven’t really found anything that hits home:
In the end I want to stack the images in a tensor on top of each other, to the dimension (WxHxDxN). Can I combine the dataset’s
__getitem__ and the sampler to make sure that the first sample contains the
N first images and the last sample containts the
N last images? My spontaneous feeling is that I should yield the tensor in
__getitem__, thus being able to easily create batches of “sequences” as the sequences are already stacked, but how can one create magic with restricting the indices?.
.@ptrblck answers in the first and second link are definitely interesting, and I’ve looked at using SequentialSampler/BatchSampler, but in this case it doesn’t solve the problem of “starting” at the Nth image.