In the context of the learn-real project, the training of robot related models requires the use of a very high number of images which can only be optimally loaded using PyTorch iterators.

This cannot be done currently as all data goes through the database paradigm which cannot be used directly by PyTorch iterators.

Using the standard BEAT approach would lead to overly complex and inefficient implementation.

The solution proposed is that a new property can be inserted in the block configuration so that the root path of the database is also mounted on the algorithm container.

This will only work on algorithm containers that are connected to a database.