Home

Welcome to the returnn_common wiki!

Starting points:

• RETURNN common homepage has an introduction and small usage examples
• RETURNN common principles
• RETURNN common conventions
• RETURNN example config, with get_network and a RETURNN-common model
• RETURNN migration guide
• Docstrings in the code. It is anyway very recommended to use an IDE to be able to use auto-completion, and the IDE would also automatically show you the documentation.
• nn.base docstring. nn.base defines many important base classes such as nn.Tensor, nn.Module, and has some high-level explanation of how it works internally
• nn.naming docstring. nn.naming defines the layer names and parameter names, i.e. how a model (via nn.Module) and all intermediate computations map to RETURNN layers.
• Missing pieces for first release and Intermediate usage before first release. This is an overview, also linking to now completed issues. The issues often come with some discussion where you find the rationality behind certain design decisions.
• Translating TF or PyTorch or other code to returnn_common