- 25 8月, 2020 40 次提交
-
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 328319683 Change-Id: I674a6c0d789058737df9adaadebed5a56dfc21c1
-
由 Mehdi Amini 提交于
We've been moving to an explicit registration mechanism PiperOrigin-RevId: 328314576 Change-Id: I41c93fe142d64e1bc43f3c0e1626c492ffbda92d
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328311938 Change-Id: I13a95c2c15d35e4ff051a9c45fc21a94c4cf38d8
-
由 Benjamin Kramer 提交于
Updates LLVM usage to match [c6fb72de4f55](https://github.com/llvm/llvm-project/commit/c6fb72de4f55) PiperOrigin-RevId: 328304428 Change-Id: I93e706dbe926e3d65ee6be498f54978097bfe8a2
-
由 Henry Tan 提交于
PiperOrigin-RevId: 328295550 Change-Id: I823a77ba5c56cf03d2aaf1ccfe090b500f73e5a0
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 328291622 Change-Id: If3aa7cfed1838bf490603828454257dd78582381
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 328291621 Change-Id: If04388ad4e881890383fa7e83b49c272ff216949
-
由 A. Unique TensorFlower 提交于
Converting the individual ops into tfl.softmax will improve performance by enabling fusion as well as imporve accuracy on backends where the intermediate tensors have reduced precision. PiperOrigin-RevId: 328287102 Change-Id: Ie2e37804c20854ba57ae7b1d11a4620ea17bd39f
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328283554 Change-Id: I6eb06cb27648ce2442c7ce5fd4f4720bb211986e
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328283394 Change-Id: I266dd2fb23195581494e36d1f9cdf9c1d3006943
-
由 Mehdi Amini 提交于
We've been moving to explicit registration for every users. PiperOrigin-RevId: 328274615 Change-Id: I3c3f83d107ff8131d441640d23b82aaeff86a0b5
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328273735 Change-Id: I57b78630da1dd4482114fdba7a2a6e2493d83b08
-
由 Mehdi Amini 提交于
Registration is explicit everywhere now. PiperOrigin-RevId: 328263769 Change-Id: Id76baa3276b160f0d9f1a5a5fe2cb9a9474c1d0e
-
由 Mehdi Amini 提交于
Remove dependency on the MLIR Global Dialect registry from third_party/tensorflow/compiler/mlir/tensorflow/... (NFC) PiperOrigin-RevId: 328256230 Change-Id: I180650c53c9bbb790bead9d47ae546a3938387d1
-
由 Natasha Kononenko 提交于
Implement FFT 1D lowering to hlo. For RFFT/IRFFT a custom lowering is implemented in order to support padding/slicing logic that matches tf2xla. PiperOrigin-RevId: 328252791 Change-Id: I2ad9ee5d13317f05ab2b24543e8f72289bb561c6
-
由 Robert David 提交于
PiperOrigin-RevId: 328250447 Change-Id: I68385873ab7eabdfb18e3f72d46cc76e2998bbc6
-
由 Zhenyu Tan 提交于
PiperOrigin-RevId: 328247968 Change-Id: Ic8e4d4dee29ebfb20cf1317f72a51b8db0bd78af
-
由 Lu Wang 提交于
PiperOrigin-RevId: 328242690 Change-Id: Icaaba6c49ed201090d8e2cfa42c7f666143e8927
-
由 Chenkai Kuang 提交于
1. Dense layer partition. It is learned that in some cases dense layer partition could improve the model training speed. 2. ShardedVariable now supports "assign", "assign_add" and "assign_sub" methods. 3. ParameterServerStrategy now accepts a "variable_partitioner" parameter that controls all variable partitioning under strategy.scope(). It is compatible with tf.compat.v1 partitioner. Default partitioner is same as estimator canned models: each partition would has at least 64MB data. 4. ParameterServerStrategy now is able to do memory-efficient initialization of sharded variables, but it requires a custom initializer that is partition aware. 5. ParameterServerStrategy now is able to partition variables even if their `initial_value` is a Tensor (not a callable). Meanwhile, removed `strategy.experimental_variable_partitioning_scope` method. Per-layer partitioning using different partitioners is not going to be supported right now. PiperOrigin-RevId: 328241842 Change-Id: I382743dd8d1a2f6b7ab207576aed2e77d71c5735
-
由 Mehdi Amini 提交于
Remove dependency on the MLIR global dialect registry from third_party/tensorflow/compiler/mlir/python/... (NFC) PiperOrigin-RevId: 328241726 Change-Id: Ife6f7d0717c39000f04ad8c95f9e34286628d801
-
由 Andy Ly 提交于
This follows how shape inference is performed for passthrough ops like tf_device.launch. PiperOrigin-RevId: 328241534 Change-Id: I83cc618291f5dc84ba18f963b413d5117e98de8c
-
由 Akshay Modi 提交于
PiperOrigin-RevId: 328241384 Change-Id: I6f7d3c5fd73f1c8b31c178c2b69a786a2a2d44e3
-
由 Advait Jain 提交于
accepting optimized kernel PRs. PiperOrigin-RevId: 328240994 Change-Id: I5b5a58ae196cc5f9672dfad2e78459f69369f717
-
由 Raman Sarokin 提交于
PiperOrigin-RevId: 328239471 Change-Id: I404f86d58a73b353c49aa56fefcba134026a1eb8
-
由 Henry Tan 提交于
PiperOrigin-RevId: 328239422 Change-Id: I57149ee2cc34de7e70a75564fdb52411499704d4
-
由 A. Unique TensorFlower 提交于
Updates LLVM usage to match [77e5a195f818](https://github.com/llvm/llvm-project/commit/77e5a195f818) PiperOrigin-RevId: 328236932 Change-Id: I7d786cccf01536487b3a4843fc132a635e62758e
-
由 Jay Shi 提交于
PiperOrigin-RevId: 328235521 Change-Id: I758bd6f18097979d28f067c78ebadc45075d98eb
-
由 Katherine Wu 提交于
PiperOrigin-RevId: 328228967 Change-Id: Ibeef04e68c553165f14ef583d28441730c688f53
-
由 Chenkai Kuang 提交于
PiperOrigin-RevId: 328228396 Change-Id: I826187eaf2105f37be83350f3b6b08b48fcf0349
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328223323 Change-Id: I4355eb517f0d849b012eb618a400edf5b85791a5
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 328222517 Change-Id: Ic5fa2faf946e8b31f492370cf291ef4c2cf7cac7
-
由 Henry Tan 提交于
PiperOrigin-RevId: 328222137 Change-Id: I1c4339867f6e887e3647f5f60c58a7cfd0885d3f
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 328216593 Change-Id: Ic8e64331dd8a9fbbf8054a2432b9b76dc8f01e41
-
由 Chen Cen 提交于
PiperOrigin-RevId: 328214869 Change-Id: I90b03017e12ab7c4280a4b10a5f823705499fbd0
-
由 Amit Patankar 提交于
PiperOrigin-RevId: 328213510 Change-Id: I01690726d594c7f67090f11aa95c9a9ff18c768d
-
由 Raman Sarokin 提交于
PiperOrigin-RevId: 328210531 Change-Id: Id64132cc7c0468ae1c50d960ed07ef1899175433
-
由 Saurabh Saxena 提交于
PiperOrigin-RevId: 328206539 Change-Id: I0b208dc20fa454acff3b892b5f0833536e62fc9f
-
由 Ken Franko 提交于
PiperOrigin-RevId: 328206256 Change-Id: I2f8a7a14d4a246fb86cd78b02e77c734e4178514
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 328206157 Change-Id: I111ed8a3bcf8266babaad7a1018fcd39495c9350
-
由 Doe Hyun Yoon 提交于
It's reported that op_level_cost_estimator crashes when input depth or output depth is zero. This CL checks any zero value in dim and just returns zero Costs. PiperOrigin-RevId: 328201526 Change-Id: I52c085d3e4b9f549a3e1d2a46c5d6b34bb696540
-