Ragged tensors for list of variable shape 2D tensors in PyTorch in order to be able to feed data of variable shape in a batch size > 1

[Mona_Jalal]

I have a list of 2D tensors in which the first dimension is varied while the second dimension is not varied.

I came across ragged tensor concept and searching for equivalents came across this answer by @ptrblck for 1D tensors of different shape.

Nested list of variable length to a tensor

This code might work: target = [[[1,2,3], [2,4,5,6]], [[1,2,3], [2,4,5,6], [2,4,6,7,8]]] max_cols = max([len(row) for batch in target for row in batch]) max_rows = max([len(batch) for batch in target]) padded = [batch + [[0] * (max_cols)] * (max_rows - len(batch)) for batch in target] padded = torch.tensor([row + [0] * (max_length - len(row)) for batch in padded for row in batch]) padded = padded.view(-1, max_rows, max_cols)

How can I extend this to a nested list of 2D tensors?
Please consider I have a batch size of 4 and my tensors are of these sizes:
[ 300x512, 1000x512, 11x512, 8000x512]

Is this what you would do when you have different sized tensors in a batch? Because when I don’t have the same size tensors, I get error and vision transformer requires me to have all of my intermediate representation of the same dimension.