Previously, we retained the concept of "default model architecture is encoder/decoder RNN", but that did not fit the modularity of the latest changes (notably when updating nested configurations from existing models).
With these changes, the model configuration key defaults to None in the training code path, which makes it necessary to pass at least an architecture when training from scratch. An explicit error is raised in that sense. (Might still be improved/better documented.)
Fix #5
Previously, we retained the concept of "default model architecture is encoder/decoder RNN", but that did not fit the modularity of the latest changes (notably when updating nested configurations from existing models).
With these changes, the
model
configuration key defaults toNone
in the training code path, which makes it necessary to pass at least an architecture when training from scratch. An explicit error is raised in that sense. (Might still be improved/better documented.)