| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 160538962
|
|
|
|
|
|
|
|
| |
per-compilation option.
Don't resolve compile time constants in while loops since we must preserve the loop signature.
PiperOrigin-RevId: 160535393
|
|
|
|
| |
PiperOrigin-RevId: 160528361
|
|
|
|
|
|
| |
and not fed. The PlaceholderOp constructor was attempting to convert its shape attribute to a TensorShape, which requires that the shape is fully defined. This failed with a misleading error message for a partial shape. Instead, use a PartialTensorShape.
PiperOrigin-RevId: 160526882
|
|
|
|
| |
PiperOrigin-RevId: 160520466
|
|
|
|
| |
PiperOrigin-RevId: 160518040
|
|
|
|
| |
PiperOrigin-RevId: 160515573
|
|
|
|
|
|
| |
of quantization bits.
PiperOrigin-RevId: 160515361
|
|
|
|
| |
PiperOrigin-RevId: 160488554
|
|
|
|
| |
PiperOrigin-RevId: 160488521
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This analysis is one of the most expensive parts of the HLO optimization
pipeline.
- Avoid one or two unnecessary hashtable lookups in
PopulateDefinedBuffersAndAliases.
- Add a mode to ShapeTree wherein we avoid copying Shapes.
- Use templated functors rather than std::function in ShapeTree's
iterators, thus avoiding the overhead of std::function.
PiperOrigin-RevId: 160487485
|
|
|
|
|
|
|
| |
This requires some plumbing in user_computation to pipe the debug options
through a few layers.
PiperOrigin-RevId: 160459822
|
|
|
|
|
|
|
| |
Also pipe debug_options through the code in hlo_graph_dumper, since the number
of individual parameters was growing too large.
PiperOrigin-RevId: 160446088
|
|
|
|
|
|
|
| |
- Change codegen_test_base and hlo_test_base to disable constant folding.
- Add convenience method ExecuteAsync to ClientLibraryTestBase.
PiperOrigin-RevId: 160440787
|
|
|
|
| |
PiperOrigin-RevId: 160436466
|
|
|
|
|
|
| |
Add main, FLAGS, and BUILD rules.
PiperOrigin-RevId: 160434643
|
|
|
|
|
|
| |
TF_CUDA_CLANG (--config=cuda / --config=cuda_clang)
PiperOrigin-RevId: 160433337
|
|
|
|
| |
PiperOrigin-RevId: 160429376
|
|
|
|
|
|
| |
transforming, and support sparse categorical.
PiperOrigin-RevId: 160424501
|
|
|
|
| |
PiperOrigin-RevId: 160423850
|
|
|
|
|
|
|
| |
The goal is to make kernels mostly independent of proto headers, which will let us lock down our .so imports.
RELNOTES: n/a
PiperOrigin-RevId: 160422647
|
|
|
|
| |
PiperOrigin-RevId: 160422523
|
|
|
|
|
|
| |
compatible with sklearn.
PiperOrigin-RevId: 160420406
|
|
|
|
|
|
|
| |
to enable the propagation of shapes for conditionals, which also include Merge
nodes.
PiperOrigin-RevId: 160417770
|
|
|
|
|
|
| |
Do some minor cleanup in the process.
PiperOrigin-RevId: 160404421
|
|
|
|
|
|
| |
Update the fix for the `debug_mnist` example to use `tf.nn.sigmoid_cross_entropy_with_logits`.
PiperOrigin-RevId: 160397398
|
|
|
|
|
|
|
| |
Add a screenshot of the tensorboard diagram to the tutorial page.
Fixes: #3044
PiperOrigin-RevId: 160397248
|
|
|
|
| |
PiperOrigin-RevId: 160390702
|
|
|
|
| |
PiperOrigin-RevId: 160382016
|
|
|
|
| |
PiperOrigin-RevId: 160372343
|
|
|
|
|
|
| |
parameter `name` if SKCompat is used internally.
PiperOrigin-RevId: 160370884
|
|
|
|
| |
PiperOrigin-RevId: 160369034
|
|
|
|
|
|
|
|
| |
(1) Account for control dependencies in reachability.
(2) Invert sense of reachability. We draw our HLO graphs with arrows from producers to consumers so it makes more sense for reachability to be defined along the direction of these edges.
(3) Rename ComputeTransitiveOperands to ComputeReachability.
PiperOrigin-RevId: 160366307
|
|
|
|
|
|
|
| |
This ensures we get a sane output when tfcompile is invoked manually
without only the bare minimum command line arguments.
PiperOrigin-RevId: 160363243
|
|
|
|
|
|
| |
worked.
PiperOrigin-RevId: 160362310
|
|
|
|
| |
PiperOrigin-RevId: 160354628
|
|
|
|
| |
PiperOrigin-RevId: 160354095
|
|
|
|
| |
PiperOrigin-RevId: 160351254
|
|
|
|
| |
PiperOrigin-RevId: 160347179
|
|
|
|
| |
PiperOrigin-RevId: 160346151
|
|
|
|
| |
PiperOrigin-RevId: 160344052
|
|
|
|
| |
PiperOrigin-RevId: 160341769
|
|
|
|
|
|
| |
implemented in tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py.
PiperOrigin-RevId: 160340994
|
|
|
|
| |
PiperOrigin-RevId: 160340888
|
|
|
|
|
|
|
|
|
| |
tf.scatter_nd_non_aliasing_add acts similarly to tf.scatter_nd_add but
works on non-ref objects (i.e., Tensors -- not Variables). This means
it has a gradient with respect to the primary input as well as the
updates. It does its best to avoid making extra copies of the input.
PiperOrigin-RevId: 160339328
|
|
|
|
| |
PiperOrigin-RevId: 160339087
|
|
|
|
| |
PiperOrigin-RevId: 160338811
|
|
|
|
| |
PiperOrigin-RevId: 160338775
|
|
|
|
| |
PiperOrigin-RevId: 160334030
|
|
|
|
|
|
|
| |
I'm trying to make core/kernels independent of protos. Currently the dtype ResourceHandle is itself a proto. After this CL, ResourceHandle is a normal C++ type which gets converted to/from ResourceHandleProto at (de)serialization time.
RELNOTES: n/a
PiperOrigin-RevId: 160329002
|