experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Update the documentation of the composite tensorflow project

Browse Source 
PiperOrigin-RevId: 344328755
Change-Id: Idf73e2eca38e23a7983a2517b4434d239e969134
This commit is contained in:
Feng Liu 2020-11-25 15:47:25 -08:00 committed by TensorFlower Gardener
parent c624153938
commit b3e77852a2
2 changed files with 15 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
tensorflow/compiler/mlir/tfr/README.md
View File

@ -7,7 +7,7 @@ composition in the authoring language.
The set of standard TF ops is currently open. New ops are defined for special
purposes but it is hard to make them work end-to-end: The op
needs to be handled separately by a several backends (tf2xla bridge, tflite
needs to be handled separately by several backends (tf2xla bridge, tflite
converter, CPU kernels, etc.). Writing shape functions and gradients for these
ops is extremely difficult. `tf.function` makes some parts of the implementation
simpler, but it introduces runtime overhead and it cannot easily be used to
@ -110,10 +110,10 @@ def _my_op_c(ins):
Utilities have been built to compile the Python composition functions down to
the backend IR. The project also provides a set of graph optimization passes to
expand the composite ops in the graph by using the input backend IR. These
passes have been added to the TF [common runtime]
(https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/common_runtime)
for graph execution and [eager_runtime]
(https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/common_runtime/eager)
passes have been added to the TF
[common runtime](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/common_runtime)
for graph execution and
[eager runtime](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/common_runtime/eager)
for eager execution.
## Compiling Op Composition
@ -140,7 +140,7 @@ gen_op_libraries(
```
More composite op definitions and usages are here included in the
[examples](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/mlir/tfr/examples)
[examples](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/tfr/examples)
directory.
### Just-In-Time (JIT) mode
@ -151,6 +151,12 @@ directory.
* `while` statement
* condition of `if` statement couldn't be a tensor
## RFC
This project is an alternative implementaion of [RFC:Standardizing composite ops in tensorflow to support efficient inference](https://github.com/tensorflow/community/blob/master/rfcs/20190610-standardizing-composite_ops.md).
This project doesn't rely on the tracing functionality provided by `tf.function`
to avoid all its pitfalls and it helps to build more general transformations in
the backends.
## Team
* Feng Liu

4
tensorflow/python/eager/def_function.py
View File

@ -1531,7 +1531,9 @@ def function(func=None,
This can either be specified as just the string name of the function or
a NameAttrList corresponding to a list of key-value attributes associated
with the function name. The name of the function will be in the 'name'
field of the NameAttrList.
field of the NameAttrList. To define a formal TF op for this function
implements, try the experimental [composite TF](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/tfr)
project.
experimental_autograph_options: Optional tuple of
`tf.autograph.experimental.Feature` values.
experimental_relax_shapes: When True, `tf.function` may generate fewer,