| Commit message (Collapse) | Author | Age |
|
|
|
| |
Change: 127504952
|
|
|
|
| |
Change: 127493261
|
|
|
|
| |
Change: 127492429
|
|
|
|
|
|
|
| |
- add Scalar,Vector,Matrix shorthands and use them in existing fns.
- rename CreateUnknownShape to UnknownShape.
- add Divide and Add functions.
Change: 127487121
|
|
|
|
|
|
|
|
|
| |
**Note**: Noam originally found this bug and wrote this solution. All credit goes to him.
The change solves the problem of overloading the master device when there
are lots of separate dynamic_rnns each on a different device all in the
same graph.
Change: 127481156
|
|
|
|
| |
Change: 127480049
|
|
|
|
| |
Change: 127478534
|
|
|
|
|
|
| |
the code in sync.
Change: 127477364
|
|
|
|
| |
Change: 127474696
|
|
|
|
| |
Change: 127474274
|
|
|
|
|
| |
dimensions of a Tensor. Convenience method.
Change: 127474063
|
|
|
|
|
|
|
|
|
| |
With CL/127111314, there can be multiple wheel files (different names, same content) generated in the pip_test/whl directory. But parameterized_docker_build.sh assumes a single file, which has caused nightly docker build failures such as:
http://ci.tensorflow.org/view/Nightly/job/nightly-docker-cpu/58/
http://ci.tensorflow.org/view/Nightly/job/nightly-docker-gpu/55/
This CL fixes that.
Change: 127473464
|
|
|
|
|
|
| |
option to allow small final batch sizes to shuffle_batch and shuffle_batch_join.
Change: 127472371
|
|
|
|
| |
Change: 127470901
|
|
|
|
|
| |
When save_checkpoints_secs is small (note that the smaller this value is, the more frequent we checkpoint), say 1 second, there is a race condition that could lead to a None latest_path and None self._latest_path_step, which further results to an error, when the file "checkpoint" is read at the same time it is being updated.
Change: 127470621
|
|
|
|
|
|
| |
third_party/tensorflow/contrib/framework/BUILD tests.
Change: 127470132
|
|
|
|
| |
Change: 127468219
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This allows more rich slicing of tensors i.e.
tf.assign(var, foo[-5:-10:-1, ..., 3:5, tf.newaxis])
This does not include lvalue support for assigning to a slice.
It also does not include advanced indexing foo[bar] where bar is
a tensor.
Fixed bug in implementation where num_elements was used
instead of dims for an optimization code path.
Also make supervisor have an easier to read error message
for dim mismatches.
Change: 127463353
|
|
|
|
| |
Change: 127460861
|
|
|
|
| |
Change: 127458134
|
|
|
|
|
|
|
|
| |
Rename CreateShapeXYZ to MakeShapeXYZ.
Rename CreateDim to MakeDim.
Rename CreateUnknownDim to UnknownDim.
Rename CreateDimForScalarInput to MakeDimForScalarInput.
Change: 127457122
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In this CL:
1. Remove export_logistic_regression_estimator.
2. Deprecate internal graph_actions.export_for_serving().
3. Issue warning when signature_fn is not supplied.
In future CLs:
4. Delete internal graph_actions.py (together with model.py).
5. Add signature_fn to callers where needed-- maybe via a helper function that performs the current inference, issuing a warning.
6. Require signature_fn argument to export_estimator.
7. Redirect internal usages of export_estimator.
Change: 127455493
|
|
|
|
| |
Change: 127453531
|
|
|
|
|
| |
Fixes: https://github.com/tensorflow/tensorflow/issues/3302
Change: 127449137
|
|
|
|
| |
Change: 127449098
|
|
|
|
|
|
|
| |
input to follow the same (strict) shape requirements of TensorFlow batch matrix
operations. Helper functions will reshape/brodcast vectors/matrices to make
them compatible.
Change: 127446489
|
|
|
|
|
| |
Added tf.Assert() to make sure the update_ops are not executed when is_training is False.
Change: 127443556
|
|
|
|
| |
Change: 127443499
|
|
|
|
| |
Change: 127441867
|
|
|
|
| |
Change: 127441610
|
|
|
|
| |
Change: 127440874
|
|
|
|
|
|
|
|
|
|
|
|
| |
a benchmark from ~91M for ~81M.
There are future cleanups possible but not yet done:
(a) Disallow direct destruction of WorkerInterface and always release
back to a cache of WorkerInterface objects.
(b) Replace channel cache with just a factory that knows how to create
a WorkerInterface and leave caching to the higher-level cache
of WorkerInterface objects.
Change: 127435209
|
|
|
|
|
|
|
|
|
| |
places in tensorflow::DirectSession and tensorflow::SimpleClientGraph.
Remove unused func_defs in ExecutorsAndKeys.
Fixes a few memory leaks on error paths.
Change: 127425821
|
|
|
|
|
|
| |
accommodate data-dependent notions of loss.
Change: 127424766
|
|
|
|
|
|
| |
integration with frameworks that identify them by name.
Change: 127387066
|
|
|
|
| |
Change: 127386123
|
|
|
|
| |
Change: 127385633
|
|
|
|
|
|
| |
with the standard io module.
Change: 127385608
|
|
|
|
| |
Change: 127385333
|
|
|
|
| |
Change: 127380299
|
|
|
|
| |
Change: 127380001
|
|
|
|
|
| |
(executing a TODO).
Change: 127378577
|
|
|
|
| |
Change: 127378533
|
|
|
|
| |
Change: 127376094
|
|
|
|
| |
Change: 127374393
|
|
|
|
| |
Change: 127374168
|
|
|
|
| |
Change: 127373128
|
|
|
|
| |
Change: 127371216
|
|
|
|
| |
Change: 127370125
|
|
|
|
| |
Change: 127366296
|