GitLab

Updates to syntaxnet, including update tensorflow sub-module, bazel requirement and fix trainer crash (#479)

* syntaxnet: Cosmetic fixes recommended by python lint.

* syntaxnet: Fix crash in parser_trainer due to inconsistency between LexiconBuilder::Compute()
	   and context.pbtxt definition ('char-map' input declaration was missing).

* syntaxnet: reduce flakiness in GraphBuilderTest.

* syntaxnet: Update tensorflow submodule to version > 0.10.

* syntaxnet: Update to latest stable bazel (0.3.1).

This update comes partially to allow Tensorflow submodule to build
succesffuly. In this commit, I also update and simplify the WORKSPACE,
to avoid declaring dependencies already present in tensorflow.

* syntaxnet: Update bazel version check to require version 0.3.0

* syntaxnet: Document pip requirement, along with python mock module.

Use open() instead of tf.gfile.FastGFile()

Close the TFRecordWriter after use.

Use tar on OSX to unzip the MSCOCO data file.

Remove flag --config=cuda. It's not necessary and can cause a warning.

Example of data conversion and fix comments

Fix signature change of resize_images()

Fix bug caused by signature change of resize_images().

Fix a bug in the im2txt code where the Saver is created before the optimizer

optimizer.

Add Bazel workspace name to fix bug in relative path of shell scripts.

Update citation in README.md

This reverts commit 091d6e4a.

This reverts commit a704458a.

Swap comments for the two conv2d ops.

Fix README

Add pre-trained lm_1b model

Open source the image-to-text model based on the "Show and Tell" paper.

with italics and subscript markers.

Refactoring of precision calculation.

Do not perform averaging of the cost.