Fix indexing in data sampling for training labels

[Konat23]

Hi,

I noticed an issue in TripleCartesianProd.train_next_batch when using tuple batch_size (i.e., separate batch sizes for branch and trunk inputs).

Currently, the code uses:

self.train_y[indices_branch, indices_trunk]

However, this performs NumPy advanced indexing over paired indices instead of selecting the Cartesian product. As a result, it returns a 1D array of shape (batch_size,) instead of the expected 2D submatrix (len(indices_branch), len(indices_trunk)).

Since TripleCartesianProd represents data in Cartesian product form (y_train with shape (N1, N2)), the batch selection should also preserve this structur

Why this matters

This affects training when using:
batch_size = (branch_batch_size, trunk_batch_size)

and may lead to:

• incorrect target shapes,
• silent broadcasting issues,
• or mismatches between model outputs and labels.

[echen5503]

Can you provide code that demonstrates this issue?