hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
ac68d4275a6a65f7132d4bbc34fee8ee97d382e1 | ci_build: improve documentation | Jan Prach | jendap@google.com | 1,454,060,370,000 |
89c925c2ac049dbc6bd796fb7686f1dc86274a38 | ci_build: mount cuda devices and libraries into container | Jan Prach | jendap@google.com | 1,454,060,624,000 |
5fb0a9f6f467b64b0acc97004f02643626817e5d | Merge pull request #922 from cjauvin/fix/udacity-notebooks-py3. | Vincent Vanhoucke | vanhoucke@google.com | 1,454,078,191,000 |
2cb25ab1a94f3fe32bc6e00144c278373e404eaa | Upgraded to a newer version of Eigen that fixes a compilation error when
compiling with nvcc and clang | Benoit Steiner | benoit.steiner.goog@gmail.com | 1,453,920,503,000 |
8a59748c087a2fee535c0d5067dbabb01920e812 | Use cc_binary rather than cc_library to reduce size of native library in APK from 5.5mb to 3.2mb (compressed).
Change: 113369407 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,088,858,000 |
ab33cd7106935bdda8f9fa298bf9bbd1877970c2 | The BUILD file in this folder exports a LICENSE file that does not exist.
Change: 113371009 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,089,873,000 |
d041dfa7adfe1a24957f1e6866286d13c3ad5b8f | Backported several changes from the upstream version of Eigen
Change: 113371678 | Benoit Steiner | benoit.steiner.goog@gmail.com | 1,454,090,307,000 |
5dc0ab7565afc0f707adc628c2865233a9702cfa | Inline read/write ops of Stack and TensorArray.
Change: 113373813 | Yuan Yu | yuanbyu@google.com | 1,454,091,513,000 |
938902bade73fb90ecc178a6c46cdf2b30d3a69a | Sped up construction of BCast helper class for the very common case
where both shapes are the same by using much more straightforward
code to achieve the same ultimate initialization of the various
instance variables with simpler code.
Added benchmark for this to bcast_test.cc. Speeds up the same_shape
case by 65% (67 ns to 23 ns for a two-dimensional shape) without
any real effect on the different shape case.
Run on machine with (40 X 2801 MHz CPUs); 2016/01/28-11:12:26
CPU: Intel Ivybridge with HyperThreading (20 cores) dL1:32KB dL2:256KB dL3:25MB
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_BCastSetup/0 122 122 +0.0%
BM_BCastSetup/1 67 23 +65.7%
Change: 113374076 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,091,654,000 |
6d925484a1defc9594934175a2b6e88baf418063 | Inline the various reference-count maintenance routines, since
most of them expand into just a handful of instructions and they
are performance critical.
Change: 113374152 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,091,692,000 |
34d5a998a02d274c645d29b350a6e51736a7c2f4 | Remove the AutoUpdate methods from EventAccumulator and EventMultiplexer.
There's no reason for those classes to own the async behavior, a caller can easily manage themselves.
The resulting code is easier to reason about because there is always one source of async and one thread rather than several different subtimers firing at different times. Also, it logs load time for performance measurement.
Change: 113376718 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,093,081,000 |
432852375ec07cde915246c841d18d3993236f17 | Lazily allocate the referenced tensors set in UniqueTensorReferences.
Since the common case is that no elements will be put into the
set, this helps with both CPU cost and reduces the cache-trashing
footprint of constructing a UniqueTensorReferences object (which is
done on every op in setting up the OpKernelContext).
Change: 113377001 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,093,226,000 |
743e80d6345f85f4410f61a064da32bc75fe7775 | Added __neg__ to IndexedSliced and scalar_mul to math_ops, with tests.
Added convert_to_tensor to values, indices and dense_shape for IndexedSliced creation.
TODO: Update math_grad.py to use scalar_mul(scale, grad) when possible.
Change: 113384053 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,097,224,000 |
f89ba4869abe8aec1f05dfd363da430a9fc37d35 | Colocate the constant initializer of the adagrad accumulator with the variable
and the accumulator.
Change: 113391069 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,101,812,000 |
e09ee007c019d77a49934a76661001bc508a663a | Add rename() function to gfile.py
Change: 113396524 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,105,370,000 |
b672b361a41a2b36810e262859acf93bed31ce0f | Add a logging module that can be configured to log to an events file.
The design here is that tensorboard_logging both logs to an event file via a
SummaryWriter and logs to the tf.logging mechanism.
Change: 113410039 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,113,887,000 |
b2a64c38085cbad2cc353738739e9d08b51cb61c | String equality op.
Change: 113410311 | Eugene Brevdo | ebrevdo@gmail.com | 1,454,114,090,000 |
6b8354971eb57a9c5dd5f855118178edc6293831 | Add a guide to the GraphDef file format for tool developers.
Change: 113413158 | Pete Warden | pete@petewarden.com | 1,454,116,238,000 |
97f585d506cccc57dc98f234f4d5fcd824dd3c03 | Merge commit for internal changes | Vijay Vasudevan | vrv@google.com | 1,454,128,279,000 |
614bf4a8231e00f3cabf9b08171ab9fcce2ce56d | Merge pull request #928 from jendap/ci_build
ci_build: support gpu build and interactive workflow better | Martin Wicke | martin.wicke@gmail.com | 1,454,133,734,000 |
320d723c9ddfea5ec225082d695b111c95ed76cd | Adding --recurse-submodules flag to git clone in ci_bulid README | Shanqing Cai | cais@google.com | 1,454,081,696,000 |
2b7d6220c024a332e548c7657a7149666cd1d535 | Merge pull request #934 from caisq/dev2. | Martin Wicke | wicke@google.com | 1,454,134,852,000 |
46944e4040a3a51ec097b9674a07ba2dfbcc5bf4 | + Multilayer RNN | terrytangyuan | terrytangyuan@gmail.com | 1,454,179,062,000 |
5348c3f056317065d4b8b1aab64c6e5fbaadbb41 | + Multilayer bidirection RNN | terrytangyuan | terrytangyuan@gmail.com | 1,454,179,135,000 |
88507bc9c73eb38e5c52979d0bea18d8befd9dd7 | Squashed commit of the following:
commit 4e7b6e2799d6a2bf52027dadffb4ce3f14ed3f66
Merge: b0b231a 2b7d622
Author: Christian Jauvin <cjauvin@gmail.com>
Date: Sat Jan 30 13:11:00 2016 -0500
Merge branch 'master' into fix/udacity-notebooks-py3
commit b0b231a7202db3d4b4e31fa361cd1078719b9ba9
Author: Christian Jauvin <cjauvin@gmail.com>
Date: Sat Jan 30 12:57:32 2016 -0500
More explicit examples of batch generation algo | Vincent Vanhoucke | vanhoucke@google.com | 1,454,190,206,000 |
5962171ce676533b69310c569354c7e3692d5df8 | Merge pull request #943 from cjauvin/fix/udacity-notebooks-py3. | Vincent Vanhoucke | vanhoucke@google.com | 1,454,190,231,000 |
79eba2c172417ef4cc7e68482b001738610accff | Delete tensorflow/core/public/{env.h,partial_tensor_shape.h} now that
references have been moved to their new location.
Change: 113448890 | Josh Levenberg | josh11b@tensorflow.org | 1,454,192,550,000 |
8f5da79619fc356565c51518138d64c9a71e27a4 | TensorFlow: fix OSS cuda build by adding header for int64 definition.
Change: 113453561 | Vijay Vasudevan | vrv@google.com | 1,454,205,709,000 |
48c0e9e4d6d60b06778293179ccd087a31f0fc09 | Merge commit for internal changes | Vijay Vasudevan | vrv@google.com | 1,454,225,745,000 |
90ab82a6013506eb0382606e391303e610669769 | Squashed commit of the following:
commit abdb18cbecee59358db925043bc6203f06c11280
Merge: 6082549 a27d844
Author: Patrick Christ <patrick.christ@ph.tum.de>
Date: Mon Jan 25 12:16:59 2016 +0100
Merge branch 'master' into master
commit 60825498e6166e97ab6245773840401d146d7d74
Author: Patrick Christ <patrick.christ@ph.tum.de>
Date: Sat Jan 23 16:44:37 2016 +0100
Add link to Tensorflow Udacity course
Add link to Tensorflow Udacity course
commit 958059040502849671fc0081fdff3d189ccfd676
Author: Patrick Christ <patrick.christ@ph.tum.de>
Date: Sat Jan 23 16:39:52 2016 +0100
Add link to Tensorflow Udacity course
Add link to the new tensorflow udacity course.
https://www.udacity.com/course/deep-learning--ud730 | Vincent Vanhoucke | vanhoucke@google.com | 1,454,256,636,000 |
7322dcef2cf1736df67c9a499938d5cd04a83b8c | Merge pull request #853 from PatrickChrist/master. | Vincent Vanhoucke | vanhoucke@google.com | 1,454,256,648,000 |
bf045a7ac6c7af370fd8014178b1afd58cc30085 | Squashed commit of the following:
commit b52273baaa13ea000cf904b4ea4e215410bca908
Merge: 8149625 48c0e9e
Author: Romain Thouvenin <romain.thouvenin@gmail.com>
Date: Sun Jan 31 09:01:47 2016 +0100
Merge branch 'master' into notmnist-lowmem
commit 81496253f4b2511cf2bbc72125a12bc8240f7565
Author: rthouvenin <romain.thouvenin@gmail.com>
Date: Sat Jan 30 10:39:02 2016 +0100
Load dataset classes separately to require less memory | Vincent Vanhoucke | vanhoucke@google.com | 1,454,258,228,000 |
c9dde106b48e74b8a9cb1dcb3a5b43f24fc229e6 | Merge pull request #940 from rthouvenin/notmnist-lowmem. | Vincent Vanhoucke | vanhoucke@google.com | 1,454,258,243,000 |
10cca0c34279fb3e8aaa55b236f8962e0a2eaf80 | typo in xw_plus_b | Taehoon Kim | carpedm20@gmail.com | 1,454,329,736,000 |
876cc42ac07f0e4f5d51c0f71d79bd59936dbbff | Merge pull request #949 from carpedm20/master. | Vijay Vasudevan | vrv@google.com | 1,454,353,118,000 |
0f516e3d86196ca2136f8e8b05bef6d93f2509c1 | fix dropout management in TensorFlowEstimator._predict - get correct list of dropout variables, pass keep_prob = 1.0 | Konstantin Lopuhin | kostia.lopuhin@gmail.com | 1,454,359,236,000 |
6c436f20a4ccd195e7f0e711384ff9763d32c0ba | Get rid of mysterious comment previously needed for syncing with
the Google-internal version.
Change: 113552310 | Josh Levenberg | josh11b@tensorflow.org | 1,454,356,119,000 |
20c7d5f8846c8457dc1d87bb739564f79a6c079d | Add public framework headers and library to pip package.
- Added a Skylark macro to collect transitive header dependencies of a dep.
- Exported just the framework components so we can do -ltensorflow_framework.
Change: 113558248 | Manjunath Kudlur | keveman@gmail.com | 1,454,359,792,000 |
f3464f87d22bd27599046639aea434dedf002afe | Merge commit for internal changes | Manjunath Kudlur | keveman@gmail.com | 1,454,360,065,000 |
05cae1f96f7ac2dfb6368fd87e6254194a04281a | Merge pull request #89 from lopuhin/fix-dropout
Fix dropout management in TensorFlowEstimator._predict | Illia Polosukhin | ilblackdragon@gmail.com | 1,454,363,219,000 |
b98a67d324476f25dd955cb951c1f3a76a5247b4 | Change types of constant numeric literals.
This PR fixes mismatched types in numeric literals usages.
For examples, use 'size_t{0}' and '1u' instead of '0' and '1' respectively.
This removes 26 simple compiler warnings in 20 files
on 'signed' and 'unsigned' comparisons on CHECK_* and DCHECK_* macros. | Dongjoon Hyun | dongjoon@apache.org | 1,454,208,504,000 |
39d7ff235d9e91e52888f74002dd7404ab7914a4 | Merge pull request #944 from dongjoon-hyun/change_types_of_constant_number_literals. | Vijay Vasudevan | vrv@google.com | 1,454,366,518,000 |
3083daff1da69797f9ff501391809e0f87060719 | Adds support for 16 bit PNG encoding and decoding.
Change: 113562498 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,362,453,000 |
ea6c9a3391f944fb27ca683ab7f1644711030f4b | Display sample data on the /scalars endpoint if no tag or run is specified.
This is intended to be used on the initial TensorBoard load. Downloading all
scalar data can be slow if there are a large number of values; this allows us to
show a coarse outline to the user in the meantime.
Change: 113569115 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,366,048,000 |
8525dbdc02b0c45c38ba16cfba26856eece2b851 | Fix for MatMul statistics transpose error, fixes #899.
Change: 113571550 | Pete Warden | pete@petewarden.com | 1,454,367,549,000 |
45bb3fdcabc519ec54184473df02c375a305174f | Remove old forwarding headers and build rules.
Change: 113575334 | Josh Levenberg | josh11b@tensorflow.org | 1,454,369,674,000 |
9bdb6fe02b8e2c08852f5279f43a7a5596034b99 | Merge commit for internal changes | Vijay Vasudevan | vrv@google.com | 1,454,370,695,000 |
bcd7679fe2c3d774819e48c369ee65d949618ff3 | fix gradient of tf.floor | Yuxin Wu | ppwwyyxxc@gmail.com | 1,454,344,917,000 |
32da216db06eaa57d422aece293ee229708dc214 | Merge pull request #951 from ppwwyyxx/tf.floor. | Vijay Vasudevan | vrv@google.com | 1,454,374,196,000 |
ce24b95483cf28226b0e5e6e7244faa19d7192e3 | * Fixed typo | terrytangyuan | terrytangyuan@gmail.com | 1,454,383,765,000 |
efe1d8f4f2c48204dbea2c5d22c7c529666c7a19 | + sequence_length for RNN model | terrytangyuan | terrytangyuan@gmail.com | 1,454,385,629,000 |
fc52290edfe3a923f68571dd14982fa69e6ca0e1 | * expose more parameter options in rnn example | terrytangyuan | terrytangyuan@gmail.com | 1,454,386,160,000 |
e882cfc139cf72a4fbaa38c52bd6467a8c71a418 | * temporarily ignore models.py coverage until finish tests for rnn | terrytangyuan | terrytangyuan@gmail.com | 1,454,387,865,000 |
3fef2bacc10d4d1c78d6a38054ff5f93c813b52d | Add the code to request the runtime permission for the camera for devices whose
API levels are >= 23. | Takeshi Hagikura | thagikura@google.com | 1,454,057,049,000 |
2edf3621622a0d2b2140a31d8030d826a82562e0 | Merge pull request #926 from thagikura/handle_runtime_permission_for_camera. | Martin Wicke | wicke@google.com | 1,454,390,994,000 |
d38f779ea5f70f63ed1a4f182ac4b591633e05aa | Adding link to Part 3 tutorial. | Illia Polosukhin | ilblackdragon@gmail.com | 1,454,396,579,000 |
4e973d2cfb5d4702461e5f6a58b497e739b22717 | removing disabled package libatlas3gf-base | Illia Polosukhin | ilblackdragon@gmail.com | 1,454,401,209,000 |
d821f6aeb66a93501673ac5314685bd7d58151f8 | Disable tensor tracking when only one GPU stream is used.
Change: 113579306 | Xiaoqiang Zheng | zhengxq@google.com | 1,454,371,852,000 |
5ff6d34a05b2eb49f7a79a4d0b78ada9a6842b6c | Fix a Python 3 incompatibility in the SWIG bindings.
Change: 113579944 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,372,231,000 |
a8d2f0983ecdea8ff2526c717d6a9b2f06f403d8 | Minor formatting fixes.
Change: 113582098 | Vijay Vasudevan | vrv@google.com | 1,454,373,556,000 |
bf20a3a53296c0cc4a2c03e2d0b9a81703b7e8d6 | Print out the Python error if an exception occurred during evaluation of a
py_func node.
Change: 113584977 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,375,377,000 |
0ea345362f1b7ffa413427157a17f5d0f4da04f2 | TensorFlow: fix EXPECT_OK and ASSERT_OK macros.
Change: 113592220 | Vijay Vasudevan | vrv@google.com | 1,454,381,440,000 |
0db132c4690a07c2d77d50b43b42c730af78ce3c | Autogenerated Change: Release TensorBoard at TAG: 11
Change: 113594352 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,384,229,000 |
bc58a40a86126bf91c92cd85f7c47eb7fe4f4ca2 | Allow PyFunc to accept inputs of length zero. This is useful if a python
function is producing data for the rest of the Tensorflow graph.
Change: 113597306 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,387,889,000 |
1c167b7debf3d51e3dfdd745e59e1267e03fd02c | Fix print formatting.
More general exclusion of files (h/t @shreyasva)
Typo (h/t @seanpavlov)
Change: 113597422 | Vincent Vanhoucke | vanhoucke@google.com | 1,454,388,054,000 |
945e114994c09cd1116feb291cb9d0a2cf9954d3 | Fix build.
Change: 113599077 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,390,078,000 |
09b41ce5cc07a15df239210600caaefafa6c7198 | In ResizeBilinearOp, reduce the number of multiplications
(mostly but not entirely replaced with additions/subtractions),
also yielding a slight improvement in numeric precision.
Change: 113617221 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,409,732,000 |
02084099cf44892c2c8c09b860e0e6b297c9f12a | Merge commit for internal changes | Manjunath Kudlur | keveman@gmail.com | 1,454,430,991,000 |
34e38936d627fae97ac53f09bb333cba8c334858 | Remove external/<repositoryName>/ prefix to includes
This prefix is invalid when using tensorflow from a remote repository. | Damien Martin-Guillerez | dmarting@google.com | 1,453,133,885,000 |
437646b1495266cd4e5058f96bb06098785f4d4e | Merge pull request #969 from damienmg/external_headers. | Martin Wicke | wicke@google.com | 1,454,452,689,000 |
2603ae927dcb29ffac99a035cb89f375c1aac990 | Add tools/python_bin_path.sh to .gitignore | Geoffrey Irving | geoffreyi@google.com | 1,454,439,978,000 |
c4882da565453b37e2e17d2b275483c63b9e925f | Add a missing srcs_version = "PY2AND3" | Geoffrey Irving | geoffreyi@google.com | 1,454,440,002,000 |
0262c665184fbb3f656591b1c912007edea9f646 | Merge pull request #966 from girving/tweaks. | Vijay Vasudevan | vrv@google.com | 1,454,459,370,000 |
1e141835fb562546426d638aeba0830892208500 | Add assert_same_float_dtype.
Change: 113646664 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,435,345,000 |
de9cb2f0bfecf37cd7a0c8ee7966fe99f7d1f432 | Add a missing `import` statement to show correct error messages.
`_googletest.py` uses `sys.stderr.write` and `sys.exit` without importing `sys`. | Dongjoon Hyun | dongjoon@apache.org | 1,454,373,515,000 |
b4d364e25faaeafeafef416a88b145ba026b4f3f | Merge pull request #954 from dongjoon-hyun/add_missing_import_in_googletest. | Vijay Vasudevan | vrv@google.com | 1,454,459,402,000 |
34afdbbcf35309416568dd96422c45aaefdf0f84 | Fix typo in code comments in cifar10.py
Change: 113652958 | Dan Smilkov | dsmilkov@gmail.com | 1,454,438,462,000 |
9ede62fccd761203c17250afe5c132fb82aa689d | Select op can now broadcast when the condition is a vector.
In this case, instead of operating element-wise it operates
row-wise on arbitrary rank tensors.
This alternate calculation happens when the cond Tensor
is a vector whose size matches the row count of the 'then'
and 'else' tensors.
GPU-enabled + tests. Will be used for intelligent update of RNN states during
dynamic with computation variable sequence lengths in a minibatch.
Change: 113659369 | Eugene Brevdo | ebrevdo@gmail.com | 1,454,441,709,000 |
e4c1411584179ca91ccaba05aea81d5a5c45fcfc | Add support for string types to PyFunc.
Change: 113661044 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,442,539,000 |
7b4a4639747affd011a6abe98b6a77015e659d4e | Add RemoveNewDefaultAttrsFromGraphDef() for preparing a GraphDef for
use by a server that doesn't have the latest version of some ops.
Also switch to generally passing OpRegistryInterfaces via const
reference instead of pointers (except when an object holds on to the
pointer beyond the single method call).
Change: 113671866 | Josh Levenberg | josh11b@tensorflow.org | 1,454,448,984,000 |
474240ded5a58e89342aed976b0cc2f75b127702 | Add tf.learn.fully_connected, tf.relu_layer, and tf.learn.relu6_layer.
Change: 113672300 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,449,224,000 |
08b09699ad5f8b57add1f83461568b647c42129a | Add boolean_mask Op. boolean_mask(tensor, mask) is equivalent to numpy's
tensor[mask].
Change: 113676874 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,451,680,000 |
3b3b06278cc7971bb273563bc917a89d49d19e7d | Merge commit for internal changes | Vijay Vasudevan | vrv@google.com | 1,454,459,663,000 |
1471cf9e398ad7262855cb8ac425cfb2033b19b0 | + initial_state and modified sequence_length | terrytangyuan | terrytangyuan@gmail.com | 1,454,462,612,000 |
862769ba88821e36b1f869b33b09d1e46e2f5066 | * Working unit test for RNN classifier | terrytangyuan | terrytangyuan@gmail.com | 1,454,467,044,000 |
6587cee84508dafb1a40c9985051d39997ceec9a | + rnn regression test | terrytangyuan | terrytangyuan@gmail.com | 1,454,468,379,000 |
b729f46b2db01b4b53e2a79b773f7269fea0bc0a | Ref #57: Moving to use class weight on logits before cross entropy. The math should work as -weight[class]*x[class] + log( sum ( exp weighted x)) | Illia Polosukhin | ilblackdragon@gmail.com | 1,454,471,131,000 |
c15732f9fc8e030872ea7234d300e76a9b5c9cd7 | Register ZerosLikeOp for all number types and bool
Fixes #952 | Martin Wicke | martin.wicke@gmail.com | 1,454,517,282,000 |
1be31a13f9007415f01d212879f29d9db1247139 | Merge pull request #978 from tensorflow/fix-952. | Vijay Vasudevan | vrv@google.com | 1,454,519,922,000 |
e830638148e203a2d9cf491e4693d35661a360d1 | Refactor the logic to apply optimization into a common module.
Change: 113692577 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,460,703,000 |
7692bdac3491c4d8f70f16351e291f17fb2bac32 | Make all the expensive steps in 1_notmnist lazy, and provide an explicit path to forcing them.
This should alleviate some of the issues with having to always rerun the beginning of the Notebook, while still providing a simple way out of accidental aborts in the middle of unzipping or pickling.
h/t @mtamburrano and @shreyasva
Change: 113693113 | Vincent Vanhoucke | vanhoucke@google.com | 1,454,461,016,000 |
5919a8330fd1b7c7ac673b293b007e0e339905f7 | Updated Tensorboard logic for determining what data should be thrown out after a Tensorflow crash and restart to use SessionLog.START proto, which should be less finicky than the previous out-of-order event.step heuristic.
Change: 113698523 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,464,955,000 |
3645876eead595b945a9cc7130954a72c5f3a545 | Fixed bad Python syntax in event_accumulator.py.
Change: 113701311 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,467,729,000 |
e27da590fec8eed886ecd1cf3d0c2575dffeaa09 | Change for making framework_function_test run on GPUs
Change: 113708473 | A. Unique TensorFlower | nobody@tensorflow.org | 1,454,476,749,000 |
e1b77eb2931c60c71bc149bb1ec13394c508162b | Fix typo (h/t Yao-Nan Chen)
Change: 113748862 | Vincent Vanhoucke | vanhoucke@google.com | 1,454,519,776,000 |
520247f8a4fb59b163a71482268c830de94be09c | Improve TensorArray: reduce thread contention, add dynamic write and size.
Change: 113749476 | Eugene Brevdo | ebrevdo@gmail.com | 1,454,520,212,000 |
4b6d54b1e49b110463623ded7e121e011aef3879 | Revert "Register ZerosLikeOp for all number types and bool"
This reverts commit c15732f9fc8e030872ea7234d300e76a9b5c9cd7. | Vijay Vasudevan | vrv@google.com | 1,454,521,920,000 |
50a86284faadb57441165b6a20b5b18e3d349372 | tf.nn.rnn(..., sequence_length) now conditionally selects states and output.
What this means is that when passing sequence_length to either rnn() or
bidirectional_rnn(), the following hold for any given row r of output and state:
if time < sequence_lengths(r):
output(r) is calculated
state(r) is calculated
if time >= sequence_lengths(r):
output(r) is zero
state(r) is carried over from sequence_lengths(r) - 1.
What this means when minibatching with sequence_length is that:
* When t < min(sequence_length), all the outputs are calculated
* When t >= max(sequence_length), all the outputs are zero (and gradients to
these outputs will not be propagated)
* When min(sequence_length) < t < max(sequence_length), the outputs are
a mix of zeros and calculated values (and gradients are conditionally
propagated to only those entries which are still calculating)
* The final state's rows correspond to:
state[r] = state calculated at row r at t = sequence_lengths(r) - 1
and gradients are properly propagated to start updating weights
accordingly for each row.
Change: 113752785 | Eugene Brevdo | ebrevdo@gmail.com | 1,454,522,311,000 |
38135ad4a83ec975f0c00011a9d64121bacf052c | Updated Docker container.
Change: 113752933 | Vincent Vanhoucke | vanhoucke@google.com | 1,454,522,394,000 |
efed7fbafd8f76dea3f82742726e35b72dc1c931 | Temporary fix to tensor_array_ops_test failures with a TODO.
Change: 113753319 | Eugene Brevdo | ebrevdo@gmail.com | 1,454,522,598,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.