Loaders

Because there are many ways to implement a neural net, there's no one way to capture them. Machine learning frameworks sometimes provide externalizeable formats, which expose a varying degree of functionality. Other times, a network is built into a software library and can only be used directly. To solve this, Dnnamo provides loaders. Loaders are bits of code that take in the location or name of a network and return a corresponding Dnnamo model object. Depending on the loader and the implementation of the model, this object might be one of several model types.

Common Loaders

RunpyLoader

RunpyLoader loads a model by running a Python function. Specifically, it imports either a file or a package and looks for the special function __dnnamo_loader__. This function should take no arguments and return a instance of a subclass of dnnamo.core.model.DnnamoModel.

Required arguments:

• identifier: The name of the Python file or module to load. If your model comes from a single file, use the file name without the .py extension (as you would if you were using the import keyword in Python normally).

Optional arguments:

• pypath: A list of paths which are temporarily added to sys.path for the purposes of loading the identifier module.

TensorFlow-Specific Loaders

FathomLoader

Fathom is a set of reference workloads written in TensorFlow. This loader imports the Fathom library and dynamically creates a Dnnamo model out of one of its workloads. Note that these are currently ImmutableModels only and do not yet support runtime information.

Required arguments:

• identifier: The name of the Fathom model to load, all lowercase.

Optional arguments:

• fathompath: A single string containing the path to where the Fathom workloads are installed.