object detection validation skips batches

[DaniAffCH]

Describe the bug

micromind/recipes/object_detection/train.py

Line 189 in b4b45c0

for i in range(len(batch["batch_idx"])):

In this line, the index iterates through the tensor length, this could cause an erroneous metric computation, especially when the number of classes is reduced.

Indeed right after, you use the i to filter the bounding boxes.

micromind/recipes/object_detection/train.py

Line 190 in b4b45c0

for b in range(len(batch_bboxes_xyxy[batch["batch_idx"] == i, :])):

For instance, consider batch["batch_idx"] to be [0, 0, 0, 10]

in such case, the bbx corresponding to 10 is never taken into account as the list length is 4.

If I'm right and I'm not missing anything I would propose changing L189 to

for i in set(batch["batch_idx"].tolist()):

avoiding redundant iterations and making sure not to leave out anything.