diff options
| author |  Dan Moldovan <mdan@google.com> | 2018-07-25 18:01:50 -0700 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2018-07-25 18:08:37 -0700 |
| commit | b24037513f12a5812a21b7ea92ff904ee9ea6cd8 (patch) | |
| tree | b747b66cbe2ef8397b885eb073f77bed483a53b6 /tensorflow/java | |
| parent | 59305b118a9ed56d733b414e8ee1a272dc66466a (diff) | |
Fix bug in parallel_walk, along with a few structural bugs that this fix revealed:
1. The conversion process was inconsistently packaging the final output into modules or lists. This CL uniformly uses a list of nodes as output from all *_to_graph functions. As a side effect, converter_testing.py asserts that the output is always a single node and extracts it, so there is no need for tests to unpack it any more. Modify the compiler to skip generating a source map by default.
2. The class converter was incorrectly saving the superclass value to the string 'object' instead of the symbol `object`.
Additional refactoring that was caught along: Simplify the source mapping code, move it to origin_info.py, add tests and additional checks. Slightly simplify the error rewriting mechanism.
PiperOrigin-RevId: 206087110
Diffstat (limited to 'tensorflow/java')
0 files changed, 0 insertions, 0 deletions