How Does MultipackBatchSampler Handle Triple-Nested Lists?

[modesta-jihye-lee]

Hello! I have a question about how the MultipackBatchSampler works. Specifically, in its generate_batches function, it seems to return a triple-nested list structure, such as:

[[[12, 135, 822, ...]], [[54, 358, 522, ...]], ...]

However, my understanding is that a BatchSampler typically returns a nested list with batches structured like this:

[[12, 135, 822, ...], [54, 358, 522, ...], ...]

This triple nesting is unexpected and doesn't align with the usual behavior of BatchSampler. Am I missing something in the implementation or intended design? Any clarification would be greatly appreciated. Thank you!

[NanoCode012]

This is from my understanding of reading the code. We pre-compute batches of mini-batch using generate_batches (resulting in 3D). Then, during __iter__, we iterate over this 3D, returning 2D batches (like BatchSampler are used to). This allows us to pre-compute the batch efficiently.