Detail about the 11 iteration during training

During the training process, each batch of image embeddings will be iteratively input to the mask decoder 11 times, and the mask decoder will output mask logits 11 times. I wonder if the mask logits of each output need to be used to calculate the loss, or is it just the last prediction to get the mask logits to calculate the loss?

The issue does not name a file, test, or entry point. Trace the training process from the mask decoder call and inspect how the 11 mask-logit outputs are consumed, then document whether every iteration or only the final prediction contributes to the loss.

Written by the indexing model from the issue text.