- 22 8月, 2020 40 次提交
-
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 327934747 Change-Id: I80eca5e53d3c3d1f3bc9996418d3713812cebd92
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 327934745 Change-Id: I5b03985ca1fe0858e9a47becb5dd6615a600da90
-
由 Hanhan Wang 提交于
Handle non-expansion and non-collapsion cases by rewriting it to two reshape ops. PiperOrigin-RevId: 327926863 Change-Id: I2b9f406d505ab69d9e25e892f75f38aa03467e1e
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 327913260 Change-Id: I53c662d45c68eb56e0b5002afd5a48119b240bd6
-
由 Reed Wanderman-Milne 提交于
Currently it fails to build PiperOrigin-RevId: 327910344 Change-Id: I1622cfe2a5196b37b6033913d37b49d0c1379a09
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 327906446 Change-Id: Ifb39e840f9a8b66a52160f971007f6672f59f679
-
由 Yanhui Liang 提交于
PiperOrigin-RevId: 327904183 Change-Id: Ib9b46931f3049db5ec9b72fe0548b81c2562ea1b
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 327902700 Change-Id: I5c4d0d2f007bc74f3165f701595f0ba0e3c5a5ba
-
由 Russell Power 提交于
PiperOrigin-RevId: 327902010 Change-Id: Ic3eb19ae45fb3ff22d8d4fa9241c92104fd01b7f
-
由 Reed Wanderman-Milne 提交于
PiperOrigin-RevId: 327900094 Change-Id: I213eb1a62d247445b7b9ea83cec6a718337711f6
-
由 Frank Chen 提交于
PiperOrigin-RevId: 327899732 Change-Id: I3888f3bcac86268a18dc957df0e4b197b9867351
-
由 Reed Wanderman-Milne 提交于
PiperOrigin-RevId: 327899317 Change-Id: I329b02411bbaf66ce32d0880edadccb440f3d4f6
-
由 Berkin Ilbeyi 提交于
PiperOrigin-RevId: 327899057 Change-Id: I602aa480c35b8734b50395d1c7e0fb621ad2d0fb
-
由 Rick Chao 提交于
PSv2: API change: Use strategy.scope where variables are created, and move failure handling context manager to module level before we settle on a module for it to live in. PiperOrigin-RevId: 327896478 Change-Id: I7c2b3583a6031282b5c549eb8b936f0dc91b10c0
-
由 Henry Tan 提交于
PiperOrigin-RevId: 327894710 Change-Id: Iac493500325f49b6a0cc5158b3b3962d022935dd
-
由 A. Unique TensorFlower 提交于
Updates LLVM usage to match [02bf5632a94d](https://github.com/llvm/llvm-project/commit/02bf5632a94d) PiperOrigin-RevId: 327894378 Change-Id: Id377dccdd14a889fa8c565c0a952972677a8709d
-
由 Yunxing Dai 提交于
- Add two side outputs in ValidateStridedSliceOp to help analyze dynamic dimensions. - Correctly set strided slice op's dynamic size if the slice size (slice end) is dynamic PiperOrigin-RevId: 327893466 Change-Id: I7a5061ed1b92006c6f1f15c71661265b5bdfec25
-
由 Kibeom Kim 提交于
Add dedicated tf.core end-to-end KPI benchmarks for the following reasons: - Most key tf.core API's execution time depends on other factors like input size, etc,... - End-to-end time is important as there are overheads that's not caught by internal timing measuring e.g. b/158246276 PiperOrigin-RevId: 327893393 Change-Id: Ic01f98d98a8edc9e19f3fad64804abe916d4aee0
-
由 Yunxing Dai 提交于
- Dynamism inference is used to decide if a value is dynamic or not. - In dynamism inference, we rewrite some instructions into boolean form. E.g., We rewrite A = constant(0) B = parameter(0) ROOT A + B into A' = constant(false) B' = constant(true) ROOT A' | B' - We also don't rewrite some instructions: E.g., A = constant(0) B = parameter(0) C = constant(0) D = parameter(0) P = C == D ROOT select(P,A,B) Into A' = constant(false) B' = constant(true) C = constant(0) D = parameter(0) P = C == D ROOT select(P,A',B') We don't rewrite P, and instructions reachable from P. - This cl fixes an issue where this two forms are mixed together: E.g., A = constant(0) B = parameter(0) P = A == B ROOT select(P,A,B) Previously the pass would fail. PiperOrigin-RevId: 327889288 Change-Id: I3dd419ca5d729bb857d3fcac8fd76d47788aa5c2
-
由 Yanhua Sun 提交于
PiperOrigin-RevId: 327888730 Change-Id: Ib16049bc731638dd2ce6ceff396fac2e2ca500a1
-
由 Reed Wanderman-Milne 提交于
I broke it in 67d15573. Before, I called a function under a `std::is_same<Device, CPUDevice>` condition which cannot be linked if Device is a GPUDevice. I would expect the function not to be generated if Device is a GPUDevice due to dead code elimination, but apparently it still is on Windows. PiperOrigin-RevId: 327887418 Change-Id: Ib97e1abf1680c75dc072850cc69c761e10ac3e1e
-
由 Yuefeng Zhou 提交于
PiperOrigin-RevId: 327886890 Change-Id: Iff29a6a3aa093cf660a0194441418be651342c72
-
由 Anjali Sridhar 提交于
Refactor DistributeVariable saveable object to extend from SaveableObject instead of ResourceVariableSaveable. This allows us move to a single type of DistributedVariable with attached policies. PiperOrigin-RevId: 327884662 Change-Id: I4f3030f4a19248dfd7e9d3281a971e637735bc5f
-
由 Pete Warden 提交于
PiperOrigin-RevId: 327883228 Change-Id: I48b7c2b11bcef6116d4208ddb5fe3205ec52ce2d
-
由 Berkin Ilbeyi 提交于
This change to HLO is needed to disambiguate cross-program-prefetches and other prefetches performed over the same HloValue. This CL is in preparation for supporting freeing cross-program-prefetched buffers after their last use. PiperOrigin-RevId: 327881463 Change-Id: Id2ea6cd543589a7d49c689d44a2631a96ee9ddeb
-
由 A. Unique TensorFlower 提交于
Backward compatible. Relevant pybind11 change: https://github.com/pybind/pybind11/pull/2291 Note: This pybind11 change makes pybind11 compatible with native Python behavior: * https://docs.python.org/3/reference/datamodel.html#object.__hash__ * A class that overrides __eq__() and does not define __hash__() will have its __hash__() implicitly set to None. PiperOrigin-RevId: 327879230 Change-Id: I8573b761ec94e8889b4eb5ac30a71ee0314e6578
-
由 A. Unique TensorFlower 提交于
Updates LLVM usage to match [5e3fd471acb7](https://github.com/llvm/llvm-project/commit/5e3fd471acb7) PiperOrigin-RevId: 327878947 Change-Id: I95107da99562016b0552ab2eca8727e6d93b4a8e
-
由 Advait Jain 提交于
PiperOrigin-RevId: 327877592 Change-Id: Id92ef2dae8c33fdd3cc218382a23126268d5ffd4
-
由 Rahul Joshi 提交于
- Fold an IfRegion with constant condition by inlining the then or else region in place of the IfRegion op. PiperOrigin-RevId: 327875351 Change-Id: Ie04e32cc7ea93ae93817ad845eb80568d7e25b35
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 327874951 Change-Id: I9fa3a2d9d01ff507752dd2f226b12b135eec70a3
-
由 Amit Patankar 提交于
Convert tensorflow/core/framework:pywrap_required_headers from a filegroup to textual headers cc_library target. PiperOrigin-RevId: 327873125 Change-Id: I2221cee258a886961b6de1b89c7b0191e233cfa6
-
由 Haoyu Zhang 提交于
This is a workaround to avoid noisy reports of PS failure from workers due to transient connection errors between them. PiperOrigin-RevId: 327872097 Change-Id: I5b1feb9cac66c78df4659dfb4b5ed81350d5678c
-
由 Chuanhao Zhuge 提交于
It's OK to fold Identity op for now, since tf_to_corert pass only supports single device placement today. When we support device placer in lowering pass, we need to check if the identity op and it's previous op are placed on the same device to decide if it is foldable. PiperOrigin-RevId: 327870203 Change-Id: If43177625fbd139fb5304a94bd54d35fa5e41417
-
由 Ken Franko 提交于
PiperOrigin-RevId: 327867890 Change-Id: I1e4543fa2e3aa2045bfde72c3f195c0db4829672
-
由 Andy Ly 提交于
This keeps the logic for suffixes appended to keys in a centralized location instead of having passes handle it when creating such ops. PiperOrigin-RevId: 327867882 Change-Id: I1f6f30486fbf29d3c0028d5996d2009f69bae24a
-
由 Blake Hechtman 提交于
PiperOrigin-RevId: 327864391 Change-Id: Id021118bc279f646ec693ec4af3f1f59cb63c38e
-
由 Yunxing Dai 提交于
Previously it just ignores dynamic dimension. PiperOrigin-RevId: 327861140 Change-Id: Icfe9a6293cc28ca2b811b1810e790f4c62e1e4a3
-
由 Ruoxin Sang 提交于
PiperOrigin-RevId: 327858513 Change-Id: I907a6b808a540fc36a19d721c7361efd2f00ff99
-
由 Cesar Crusius 提交于
The experimental SAVE_VARIABLE_DEVICES currently only has an effect on the saved GraphDef. This change makes it also work with the object graph, and the experimental TF2 C++ loading infrastructure. PiperOrigin-RevId: 327852672 Change-Id: Ief106696fde51eb834d6fe9d67e18cf3c4d201b2
-
由 Berkin Ilbeyi 提交于
This is in preparation to allow cross-program prefetch buffers to be freed. PiperOrigin-RevId: 327852436 Change-Id: I78a5b8e7eb195f326867a6bd65f850d81ddba7a8
-