| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
As an optimization, replace consecutive broadcast instructions with a single equivalent broadcast in algebraic simplification.
PiperOrigin-RevId: 190127730
|
|
|
|
| |
PiperOrigin-RevId: 190125356
|
|
|
|
| |
PiperOrigin-RevId: 190123773
|
|
|
|
| |
PiperOrigin-RevId: 190123191
|
|
|
|
|
|
|
|
|
|
|
| |
If _USE_C_API = True, this change makes us always fetch shapes using
the C API after calling TF_ImportGraphDef, even if _USE_C_SHAPES =
False. This is necessary to preserve the shapes specified by the
"_output_shapes" attr on imported NodeDefs (note that this attr isn't
present on the NodeDefs of the imported nodes, so there's no other way
to recover this information after calling TF_ImportGraphDef).
PiperOrigin-RevId: 190122991
|
|
|
|
| |
PiperOrigin-RevId: 190121386
|
|
|
|
|
|
|
| |
both generated op and its python wrapper have tf_export decorator.
create_python_api.py now checks that we don't export different symbols with same name. Also, simplified some logic.
PiperOrigin-RevId: 190120505
|
|
|
|
|
|
| |
graph as a serialized GraphDef in an attr, put it into the function library.
PiperOrigin-RevId: 190118116
|
|
|
|
|
|
|
| |
properly compare the results of the original graph against that of the hand
optimized graph.
PiperOrigin-RevId: 190115606
|
|
|
|
|
|
|
|
| |
- the shape of the array
- the hardcoding of the values of the array as a single repeated constant
scalar value, turning an activations array into a constant array.
PiperOrigin-RevId: 190115218
|
|
|
|
|
|
| |
tensorflow/contrib/eager/python/examples/spinn:spinn_test under py3
PiperOrigin-RevId: 190112748
|
|
|
|
| |
PiperOrigin-RevId: 190110866
|
|
|
|
|
|
| |
Comparable to np.argsort.
PiperOrigin-RevId: 190109968
|
|
|
|
| |
PiperOrigin-RevId: 190108434
|
|
|
|
|
|
| |
by the implementations.
PiperOrigin-RevId: 190102851
|
|
|
|
| |
PiperOrigin-RevId: 190102805
|
|
|
|
| |
PiperOrigin-RevId: 190098277
|
|
|
|
|
|
|
|
|
| |
ScopedAllocatorOp allocates a large backing tensor whose slices may be
concatenated or splitted with ScopedAllocatorConcatOp and
ScopedAllocatorSplitOp respectively. These ops should only be added via
Grappler optimizations on the dataflow graph provided by the user.
PiperOrigin-RevId: 190097586
|
|
|
|
| |
PiperOrigin-RevId: 190097557
|
|
|
|
| |
PiperOrigin-RevId: 190097168
|
|
|
|
|
|
|
| |
Same behavior as tf.gradients() for graphs.
Some discussion of this choice in #783
PiperOrigin-RevId: 190096919
|
|
|
|
| |
PiperOrigin-RevId: 190094883
|
|
|
|
| |
PiperOrigin-RevId: 190087074
|
|
|
|
|
|
|
|
| |
- Resolve inconsistency in return values (pointer to start vs end of buffer)
- Instead, return the number of chars written as this turns out to be most useful to callers
- Removes the need for redundant strlen calls.
PiperOrigin-RevId: 190085812
|
|
|
|
| |
PiperOrigin-RevId: 190085757
|
|
|
|
|
|
| |
It gets flaky timeouts.
PiperOrigin-RevId: 190084932
|
|
|
|
|
|
| |
function calls and dicts in name replacements.
PiperOrigin-RevId: 190083700
|
|
|
|
| |
PiperOrigin-RevId: 190078959
|
|
|
|
|
|
| |
This got broken in 504d103a405654f029e8902d97d4dd8f3aa07513
PiperOrigin-RevId: 190077360
|
|
|
|
| |
PiperOrigin-RevId: 190074445
|
|
|
|
| |
PiperOrigin-RevId: 190071400
|
|
|
|
| |
PiperOrigin-RevId: 190067548
|
|
|
|
|
|
| |
If the contents of a pair of back-ticks match a public api symbol name insert a link.
PiperOrigin-RevId: 190051941
|
|
|
|
| |
PiperOrigin-RevId: 190051589
|
|
|
|
|
|
|
|
|
|
|
| |
Previously we only allowed to download clang when doing GPU builds.
The added skylark files use bazel's autoconf scripts, which were only added in
0.10.0. To provide nice error message for older versions of bazel
(i.e. 'version is less than 0.10' vs 'can't load @bazel_tools/cpp/...'), we
move the bazel version check into WORKSPACE file from workspace.bzl.
PiperOrigin-RevId: 190050798
|
|
|
|
| |
PiperOrigin-RevId: 190047495
|
|
|
|
| |
PiperOrigin-RevId: 190028714
|
|
|
|
| |
PiperOrigin-RevId: 190027163
|
|
|
|
| |
PiperOrigin-RevId: 190027161
|
|
|
|
| |
PiperOrigin-RevId: 190021164
|
|
|
|
| |
PiperOrigin-RevId: 190020572
|
|
|
|
|
|
| |
rewrite functionality out of it.
PiperOrigin-RevId: 190016936
|
|
|
|
| |
PiperOrigin-RevId: 190005998
|
|
|
|
| |
PiperOrigin-RevId: 190001737
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, if the .cache() iterator is not fully consumed before
being repeated, it will cause an exception to be raised to Python.
Instead, cache should act as an identity transformation and log
an error, as this will not affect the correctness of the user's
program (at the cost of an unexpected performance cost: i.e. not
actually caching).
PiperOrigin-RevId: 189999552
|
|
|
|
|
|
| |
debugging purposes.
PiperOrigin-RevId: 189997099
|
|
|
|
| |
PiperOrigin-RevId: 189996801
|
|
|
|
| |
PiperOrigin-RevId: 189983460
|
|
|
|
|
|
|
| |
This avoids potential overflow in the denominator, also makes sure accumulation is done
in high precision.
PiperOrigin-RevId: 189982655
|
|
|
|
|
|
|
|
| |
bfloat16 is a storage format, not a computation format. Doing reductions in
this reduced precision is prone to quickly overflow. Instead, emit a float32
computation, and wrap the reduce params and result in conversions to and from float32.
PiperOrigin-RevId: 189977590
|