diff options
author | Skye Wanderman-Milne <skyewm@google.com> | 2017-05-10 11:39:52 -0700 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2017-05-10 17:41:08 -0700 |
commit | b9cdcbe3a019f09eee7e7c9c9039e4533931e1e7 (patch) | |
tree | 9c6c1689c9a3eab2cecc8aaa54790c23c5a3345c /third_party/common.bzl | |
parent | 9472ba05620332fb274e643e08228b5c530e3e52 (diff) |
Create a TF_Graph alongside the Python graph.
This is a first step towards porting the Python API to use the C API. As the Python Graph and Operations are constructed, an analogous TF_Graph and TF_Operations are created via SWIG. Currently nothing is done with the TF_Graph; a next step will be switching to the new TF_Session API which runs a TF_Graph directly (instead of a GraphDef).
This new functionality is disabled by default and can be manually enabled by setting the _USE_C_API global in ops.py. For this patch I only enabled it for a single test file. I tried enabling it for all TF Python tests and manually disabling it for unsupported tests, but there were too many failing tests (although most tests passed). See ops.py for a TODO list of unsupported functionality.
I benchmarked building an Inception model, and building the TF_Graph incurs a 20% overhead to the total graph construction time. Note that this patch does not remove any existing Python functionality; another next step will be recovering this time by removing redundant Python code. There is no measurable overhead with the new functionality disabled.
PiperOrigin-RevId: 155655064
Diffstat (limited to 'third_party/common.bzl')
0 files changed, 0 insertions, 0 deletions