| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Improved support for lookup tables. These tables are often initialized from large files stored on disk, and can take tens of minutes to load: increased the initialization timeout to give grappler enough time to load them.
Change: 154303608
* Add eager (static) checking to assert_equal when constant values are available.
This raises errors during graph creation, when possible, instead of at runtime.
Change: 154304392
* Add tf.log_sigmoid
This is a numerically stable version of tf.log(tf.sigmoid(x)). It's just
-tf.nn.softplus(-x), but it's easy to add and the identity is easy to mistype.
RELNOTES: Add tf.log_sigmoid(x) = tf.log(tf.sigmoid(x)) = -tf.nn.softplus(-x).
Fixes #3719.
Change: 154308666
* GrapplerItem is a struct and not a class
Change: 154319534
* Migrate tf_storage and tf_globals to d3v4.
Change: 154319656
* Allow SVStepCounter to measure progress for other counters beyond global_step.
Change: 154326389
* Provide a user-defined default constructor to grappler::ItemConfig to avoid compiler error
Change: 154331360
* [tf contrib seq2seq] Update BeamSearchDecoder + AttentionWrapper API:
* BeamSearchDecoder no longer tries to pass a tiling operator to the cell.
* Instead the cell must "know" about the fact that batch_size is now bigger.
* Much more explicit and careful error messages if batch sizes don't match.
* Added helper function to tile the batch by beam_width.
* AttentionWrapper now accepts an initial_cell_state argument (it'll pass the value back through when zero_state() is called).
Change: 154333600
* AllocationRegistry: only check fail if two different allocator types
are defined for the same name and priority.
Change: 154333776
* Add a conversion from xla session module to a list of node structure as a step of xla computation transformation
Change: 154337650
* tfdbg: Improve the way uninitialized tensor values and DT_RESOURCE tensor values are displayed
1. Fix the bug wherein DT_RESOURCE tensors are displayed as "Uninitialized tensor", even if they are initialized.
2. For truly uninitialized tensors, provide more information by printing the TensorProto, which contains the dtype and shape.
Change: 154338128
|
A. Unique TensorFlower
Shanqing Cai
Martin Wicke