new feature for train with single class (BCE loss) and Validation score calculation on master node when distribution

[klae01]

These are features that have not been implemented yet, but are expected to be implemented in the very near future.

1. A training loop using binary cross entropy in a single class.

   • Necessity: It can work even when there is only one class.
   • Personal reason: train.py can train the regression model.
   • I hope to be able to train on data with continuous characteristics, such as measuring the water level of a river with a single picture of the river. In this case, user have to provide custom dataset.

2. Calculation of validation metric in the main node at once

   • Currently validation acc / loss are computed on each node, but I would like to be able to compute it on the main node for scalability.
   • For example, if the roc auc metric is computed and averaged over individual nodes, it can be computed larger than the actual metric.

Applying this implementation, which branch of the repository is it correct to send the pull request to?
And if I provide these two implementations, would you be willing to add them to this repository?