This is a purely mechanical change. All that is done is:
* Deleted python/keras/mixed_precision/experimental/__init__.py
* All other files in python/keras/mixed_precision/experimental/ are moved one directly up, out of the experimental/ folder
* All Python imports, BUILD dependencies, and other references to the old experimental files are adjusted to refer to the new location
This changes the API golden files, but there is no API change. The golden files referred to the full paths of the classes in "is_instance" sections, and the full paths have changed.
PiperOrigin-RevId: 338345459
Change-Id: I9eefc2bea49b71f26ef7ec3563364a3f1d54abe6
This change exports the following class symbols, and adds relevant documentation and example code to
tf.distribute.experimental.ParameterServerStrategy
tf.distribute.experimental.coordinator.ClusterCoordinator
tf.distribute.experimental.coordinator.PerWorkerValues
tf.distribute.experimental.coordinator.RemoteValue
PiperOrigin-RevId: 338151262
Change-Id: If2d1c513d30a999c728cecc2e73b75adda1948c2
While trying to build a modular file system from outside of TF code tree
(with pip install tf-nightly headers), the following error happens:
```
Symbol not found: _TF_SetStatus\n Referenced
```
The reason for the error was caused by `filesystem:filesystem_interface`
dependency in bazel is not part of the .so.
This PR add `filesystem:filesystem_interface` as part of the dependency.
Also `filesystem:filesystem_interface` replaces `filesystem:modular_filesystem`
as dependency can be reduced to just need `filesystem:filesystem_interface`.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
This PR tries to expose modular file system header in pip install package
so that downstream packages could use the header files with pip install
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
It allows RaggedTensor.values to have types other than just RaggedTensor and Tensor.
Based on draft implementation by edloper@.
PiperOrigin-RevId: 329733234
Change-Id: Iad61506d34a72e97a9611008cdc20258e1f2bd8e
- Use of the cpuinfo thirdparty library to perform CPU feature detection, including detection of dot-product instructions on iOS and detection of cache sizes. Combined with the ARM64 build change below, this means that iPhone11/SE devices now use dot-product instructions. We measured 250 Gop/s on 1 Lightning core in iPhone 11 Pro.
- ARM32: @lissyx's fix in google/ruy#69
- ARM64: build the kNeonDotprod path unconditionally (used to be linux-only).
- x86-64: enabling AVX512 instructions in the general opensource build.
PiperOrigin-RevId: 315467315
Change-Id: I0788ddc68da8d0aa2eee71f820fb5d8c188d35b2
- ARM32: @lissyx's fix in google/ruy#69
- ARM64: enabling dotprod instructions outside of Linux.
- x86-64: enabling AVX512 instructions in the general opensource build. A previous version of this CL was rolled back due to lack of support for -march=skylake-avx512 in some toolchains. Hopefully this will fare better thanks to b68dcd8713 .
PiperOrigin-RevId: 314766242
Change-Id: I56027a328248f2d85c9c74d3d8e27328ac25f332
- ARM32: @lissyx's fix in google/ruy#69
- ARM64: enabling dotprod instructions outside of Linux.
- x86-64: enabling AVX512 instructions in the general opensource build. A previous version of this CL was rolled back due to lack of support for -march=skylake-avx512 in some toolchains. Hopefully this will fare better thanks to b68dcd8713 .
PiperOrigin-RevId: 314597853
Change-Id: Ia64d5aa5bb9e47140e084e0896a5685cac247886
This PR adds mlir_graph_optimization_pass.h header to tf-nightly pip wheel.
mlir_graph_optimization_pass.h is a header file that allows to register
mlir based graph optimizaton (either part of the tensorflow, or externally
registered).
However, it is not part of the pip install so it is not possible
to register with installed version of tensorflow. This PR adds the header
file to be part of the pip install.
This PR is related to 39231
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
While trying to following the example of MlirPassthroughOp in the docstring
(https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/api_def/base_api/api_def_MlirPassthroughOp.pbtxt):
```python
import tensorflow as tf
from tensorflow.compiler.mlir.tensorflow.gen_mlir_passthrough_op import mlir_passthrough_op
mlir_module = '''python
func @main(%arg0 : tensor<10xf32>, %arg1 : tensor<10xf32>) -> tensor<10x10xf32> {
%add = "magic.op"(%arg0, %arg1) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10x10xf32>
return %ret : tensor<10x10xf32>
}
'''
@tf.function
def foo(x, y):
return mlir_passthrough_op([x, y], mlir_module, Toutputs=[tf.float32])
graph_def = foo.get_concrete_function(tf.TensorSpec([10], tf.float32), tf.TensorSpec([10], tf.float32)).graph.as_graph_def()
```
The following error occurs in tf-nightly:
```
>>> from tensorflow.compiler.mlir.tensorflow.gen_mlir_passthrough_op import mlir_passthrough_op
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'tensorflow.compiler.mlir'
>>>
>>>
```
The reason was that dependency `"//tensorflow/compiler/mlir/tensorflow:gen_mlir_passthrough_op_py"`
was not included in pip.
This PR fixes the issue.
This PR is more or less related to 38894, which fixes a typo in docstring itself.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
Instead, we create a VariableWatcher object that keeps track of variables that
have been accessed.
PiperOrigin-RevId: 307157027
Change-Id: Ifd628b421dc725ad2366af2f6f63cf52dd1511e9
tf_python_pybind_extension requires that it be used under the tensorflow/python directory.
PiperOrigin-RevId: 304490055
Change-Id: If9e16aa20e4223149f1d1aa7d732db7bae033461
While at it, expose the associated header files from tensorflow/c/ in pip package.
Note, we expose the subset of C API that doesn't require tensorflow/cc linkage;
specifically the core operations that exclude building while loops and gradient ops,
and also excluding the experimental API.
The experimental API can also be added in the future, by factoring it into
"core" and "non-core" targets. Similarly for the C eager API.
PiperOrigin-RevId: 301601988
Change-Id: I97eac79e684fc42ce90e67ee901cdcf6f7e91cbe
While at it, expose the associated header files from tensorflow/c/ in pip package.
Note, we expose the subset of C API that doesn't require tensorflow/cc linkage;
specifically the core operations that exclude building while loops and gradient ops,
and also excluding the experimental API.
The experimental API can also be added in the future, by factoring it into
"core" and "non-core" targets. Similarly for the C eager API.
PiperOrigin-RevId: 301430667
Change-Id: I5ae7f3cedfe9dc72184d39ef1147193450c3d92e
