Potential dataset selection issue in `build_train_loader`

[ChristopherKotthoff]

Description

I suspect that in the build_train_loader method, the line for i, datas in enumerate(DatasetCatalog.get(cfg.DATASETS.TRAIN[0])): always references the first dataset in cfg.DATASETS.TRAIN. If training with multiple datasets, this might lead to incorrect size values for datasets beyond the first one.

Steps to Confirm

1. Set cfg.DATASETS.TRAIN with multiple datasets.
2. Call build_train_loader with a dataset that isn't the first in the list.
3. Check if the size is based on the current dataset or always the first one.

[ChristopherKotthoff]

Problem confirmed by changing order of datasets and training a model from the ground up. Order can be seen in the naming of the model in the lower left.