Design Doc: Tensorflow as a backend

Design doc has moved to Google Docs

This task list remains in the wiki.

Tasks

• Benchmark possible optimization strategies.

  • Measure overhead of Tensorflow on small computations.

  • Compare performance of Tensorflow vs Eigen for various-sized linear algebra.

  • Measure how well Tensorflow parallelizes over independent parts of a graph.

    • Compile multiple expressions into a single graph.

• Compile and link against the Tensorflow library.

  • PR Add dual API support to Tensorflow (see `v1.1.0-nimble branch).

  • PR Support compiling against tensorflow/c/c_api.h and linking against _pywrap_tensorflow_internal.so in Nimble.

    • Add base64 decoding library to Nimble.

    • Add helpers to check TF_Status, e.g. TF_CHECK_OK().

    • Convert between NimArr<> and TF_Tensor.

    • Add helpers to handle sessions and running.

    • Optimize to share/cache sessions or graphs.

• Generate Tensorflow code.

  • PR Factor eigenization logic out of sizeProcessing.

  • branch Add an alternate Tensorflow compiler stage gated by nimbleOptions(useTensorflow = TRUE).

  • Decide what generated code should look like.

    • notebook For RCfunctions.

    • notebook For nimbleModels.

    • For inference algorithms against models.