TensorFlow is one of the most popular open source project in Github and certainly the fastest growing deep learning framework in the market. Like many other popular open source projects, there is a market for bringing distributions of those technologies to the enterprise. In the case of TensorFlow, the possibility of an enterprise distribution is even more appealing considering the importance of deep learning and artificial intelligence(AI) in modern business solutions.
Despite its robust deep learning capabilities, TensorFlow is still missing a lot of relevant features to enable its mainstream adoption in the enterprise. Fundamentally, an enterprise distribution of TensorFlow will need to leverage other open source frameworks in order to in relevant areas of enterprise AI solutions such as security, management, monitoring and others. How would a TensorFlow enterprise distribution look like? Here are a few ideas to consider:
Some Components of a TensorFlow Enterprise Distribution
1 — TensorFlow Cluster Manager: TensorFlow Clusters are groups of jobs focused on a single objective [ex: training a neural network]. More robust tools for managing and configuring clusters will be required to adopt TensorFlow in mission critical enterprise solutions.
2 — Training Tools: TensorFlow has an incredibly robust model for training models. That flexibility should be complemented with tools in order to allow non-developers to train TensorFlow programs.
3 — TensorBoard Extensions: TensorBoard is a fantastic tool for evaluating and debugging TensorFlow programs. Creating new extensions for TensorBoard will be a very welcoming addition to an enterprise distribution of TensorFlow.
4 — TensorFlow Model APIs: TensorFlow relies on the Master-Worker Service model to enable the distributed execution of deep learning graph[hs. An enterprise distribution of TensorFlow should extend this model and expose APIs directly from TensorFlow graphs in order to increase the interoperability with third party applications. Cloud machine learning(ML) stacks such as Azure ML or AWS ML enable this capability very effectively.
5 — TensorFlow Graph Store: TensorFlow graphs are persistent by design. Enterprise TensorFlow solutions could benefit from extensions to this model that leverage robust database as the persistent store for deep learning programs.
6 — Data Source Connectors: An enterprise distribution of TensorFlow should provide connectivity with mainstream enterprise databases and back-office systems in order to simplify the access to data relevant to deep learning graphs.
7 — Security Services: Services that enable security capability such as authentication, access control or data privacy will be essential in order to streamline the adoption of TensorFlow in the enterprise.
Who will Provide a TensorFlow Enterprise Distribution ?
There are several players in the data intelligence market well positioned to provide an enterprise distribution of TensorFlow. Here are my top picks:
1 — Big Data Platform Vendors: Platforms such as Cloudera, Hortonworks, MapR or DataStax can benefit from extending its capabilities into the AI space. Providing an enterprise distribution of TensorFlow that integrates with their big data platforms seems like a seamless way to make that transition.
2 — Startups: We can always count on new startups that will extend TensorFlow capabilities into the enterprise.
3 — Google: Google hasn’t been exactly successful commercializing on-premise enterprise software. However, an enterprise distribution of TensorFlow that complements Google Cloud ML stack could be an interesting idea for the internet giant.