In the past few years quite a few DL frameworks have been developed such as Caffe, TensorFlow, MXNet, Torch, and CNTK. At the same time, some DL developers have published pre-trained models targeting these frameworks. The objective of MLModelScope is to make it easier to use DL frameworks and their corresponding models. DL developers publish their framework and models through MLModelScope, a user can then browse through the different frameworks and models. Different frameworks and models exhibit different design points — some being more flexible and/or more low level than others. At the user interface level, MLModelScope hides these complexity providing a more holistic view of frameworks.

CarML is built from modular components and is designed to be extensible and customizable. Users can disable compo- nents, such as tracing, with a runtime option or conditional compilation, for example. Users can extend CarML by adding models, frameworks, or tracing hooks.

This section describes how to extend and customize MLModelScope from the following aspects: