Chapter 4 - Essential Tweaks

As I was working through, I realised the API still did not look much like the Python API. Lots of enhancements can be made to the existing API.

One of the important parts of the framework I worked on were the Ops, after all they constitute the elements of the graph. The following changes were made to Ops handling by rtensorflow :

• Unique Name Generators : I made a unique string generator for reference to all the ops created in the working session. A permutation of 5 characters allowed for a large number of unique name possibilities. I used the characteristic Op type name as the base for easier debugging for users.
• Optional Name : Using the string generator now allowed for users to give optional custom names for the nodes created. This is useful while exporting graphs.
• Generic Routines : This development made adding new Ops to the package more streamlined. I created four basic deeper level functions for creating a Placeholder, Constant, Unary Ops and Binary Ops. This allowed for minimal wrappers in R to function as entry points to adding new Ops to the package.
• Constant Op Tensor Shape : To emulate the Python API, the constant op now does not require an explicit shape of Tensor value to be provided.

The output from the Session is now returned to the R workspace itself as a multidimensional matrix. Then I removed the explicit function for fetching outputs (fetchOutputs()). The output op is specified while running the session. The output also is returned by the same function to the R workspace.

I used templates for better data type compatibility, now the package supports float, int32, double, and boolean. Adding the remaining types is a low hanging fruit.

I also fixed a bug which did not allow multiple feeds to be passed to placeholders. It was a matter of clearing input values in wrong points of the workflow.

A crucial component for a complete R package was still missing, Documentation! I updated Documentation for all the exposed functions. Still need to add comments for utility functions deep inside the package. I updated README for easy installation and set up for new users.

Then for ease of new contributors, I refactored the whole code base following the Google C++ style guide.

My next plan is to add a Vignette for an introduction to the package, work on adding important ops (especially activation functions), and the interesting enhancement,loading saved models from tensorflow.