| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Update inception score to match the openAI version from https://github.com/openai/improved-gan/tree/master/inception_score.
PiperOrigin-RevId: 172562573
* Stub support for retrieving LossFunction by name.
PiperOrigin-RevId: 172588516
* Don't emit fusion computations separately in HloModule::ToString. These computations are emitted with their fusion instruction and therefore don't need to be emitted as a separate comptutation in the module.
PiperOrigin-RevId: 172612725
* Make `tf.contrib.distributions` quadrature family parameterized by
`quadrature_grid_and_prob` vs `quadrature_degree`. Enables support of
quadrature methods other than Gauss-Hermite.
PiperOrigin-RevId: 172622919
* Fixes test breakage.
PiperOrigin-RevId: 172626499
* Remove global step read dependency from model_fn. Estimator behavior still will be deterministic since the step checking logic in session_run_hooks was changed as follows:
* assume stale step
* before using the step, check for the current value by session.run
PiperOrigin-RevId: 172629797
* More changs to avoid flakes in random_shuffle_queue_test
PiperOrigin-RevId: 172630989
* Add expected keys to predictor exception if unexpected key detected.
PiperOrigin-RevId: 172634275
* Add TF_GraphGetOpDef() to C API and use in Operation.op_def()
Note that this creates a small change in behavior with the C API
enabled, since previously not all Python Operations had an OpDef
(op_def() returns None). With the C API enabled, op_def() always
returns an OpDef.
PiperOrigin-RevId: 172634411
* Implement ZlibInputStream::Tell() by keeping track of the number of bytes
consumed by the reader.
PiperOrigin-RevId: 172634455
* Upgrade tensorflow pip dependency version to 3.4.0+
PiperOrigin-RevId: 172635727
* [XLA] Deterministically dump an executable.
Previously, dumping a executable is nondeterministic as a map in protobuf is serialized in random order.
This CL enables "Deterministic dump" mode of protobuf, which sorts the map first before dumping them. This is helpful in comparing if two dumps are the same in XLA determinism test.
PiperOrigin-RevId: 172637100
* Fixed work size computation in Split and SplitV ops to avoid integer overflow.
PiperOrigin-RevId: 172637818
* Internal change.
PiperOrigin-RevId: 172641543
* Bug fixes for fold_constants_lib.
1. Tensor names in TF may be in the form of "a:0", "a:1", or "a" as a shorthand
notation of "a:0". FoldConstant library always expected the shorthand notation,
and did not handle the cases where explicit notation was passed to input or
output list. This means that this library could not handle the case when input
or output were not the first output of a node.
2. To match the input nodes in the original graph and the added Recv nodes in
rewritten graph, FoldConstant library used prefix matching. Unfortunately, this
means that when a input name is a prefix of another input name, there is
possibility that wrong Recv node gets matched. For example, if input names were
"placeholder" and "placeholder_1", then it did not handle the case very well.
3. RemoveUnusedNodes() in FoldConstants lib could remove nodes which output
depended on. This happened when an input name points to a node with multiple
outputs and not all outputs of that node were included in the input names.
4. ReplaceSendRecvs() in FoldConstants lib assumed that all input nodes are
removed during rewriting the graph. This assumption is not necessarily true,
and it could add a duplicate node in the graph.
PiperOrigin-RevId: 172641947
* Adds visibility to sgdr_learning_rate_decay.
Currently SGD with warm restarts is siloed in
tensorflow/contrib/training/python/training/sgdr_learning_rate_decay.py, since
it is not listed in the 'training_py' build filegroup. This change simply adds
sgdr_learning_rate_decay to this filegroup so that other projects can use warm
restarts during optimization.
PiperOrigin-RevId: 172643218
* Add logging verbosity to mnist.py
PiperOrigin-RevId: 172643922
* Automated g4 rollback of changelist 172336111
PiperOrigin-RevId: 172645893
* Correct the docstring to reflect that the values of cols_to_vars are always lists of Variable's (never single Variable's or PartitionedVariables), and make this true for bias.
PiperOrigin-RevId: 172646456
* Changes MultiLabelHead.create_loss to return a Tensor of size [batch_size, 1], to be consistent with other heads.
PiperOrigin-RevId: 172647355
* Disabling failing contrib tests.
|
Amit Patankar
Saurabh Saxena
Derek Murray
Rohan Jain
Vijay Vasudevan
A. Unique TensorFlower