hash
stringlengths
40
40
msg
stringlengths
1
131k
author
stringlengths
1
33
email
stringlengths
0
57
date
int64
1,447B
1,698B
0269c3690b402c63100869cbb7f8a8ea190e8c0a
Added export_meta_graph() and import_meta_graph() for serializing/de-serializing the graph and other Python objects necessary restarting training, running eval, or running inference into a MetaGraphDef protocol buffer. MetaGraphDef contains the following: - MetaInfoDef: For storing version and other meta data associated with the meta graph. - GraphDef: The Graph. - SaverDef: The Saver. - CollectionDef * Int64List * FloatList * BytesList * NodeList * AnyList These are evolving APIs and subject to change. Change: 114026857
A. Unique TensorFlower
nobody@tensorflow.org
1,454,777,600,000
3ab95b1f76e3edba269aaed850d797785e309230
The bug was in creating ZerosLike tensors when the shape is unknown. Change: 114029555
Yuan Yu
yuanbyu@google.com
1,454,784,587,000
1115799825078f2037d60bc2dc0283bb40f309f4
Add weighted moving average Op. Change: 114039411
A. Unique TensorFlower
nobody@tensorflow.org
1,454,811,147,000
0a00cd3b41fe1c3632a6062e11936e7bc4b3308e
Marks a few stateful ops stateful. Change: 114039491
A. Unique TensorFlower
nobody@tensorflow.org
1,454,811,433,000
aa976ef6a20865b41ab2521756d2af4d8ebb6d5a
Fix python3 issues Change: 114043720
Martin Wicke
wicke@google.com
1,454,822,806,000
23eccfb17635bce1c19b668986dceae1281ccee8
Adds c++ grad for ExpandDims and Transpose. Change: 114043957
A. Unique TensorFlower
nobody@tensorflow.org
1,454,823,505,000
dffd6457e49d2ee7e1074ffe227fa858c188ec74
Fix OSS build reference Change: 114044938
Vijay Vasudevan
vrv@google.com
1,454,825,949,000
ee3969482714615757e170d12629f721c5830536
Merge commit for internal changes
Vijay Vasudevan
vrv@google.com
1,454,826,123,000
d53f06b6a3c77160ad955b718a04260674752298
https://github.com/tensorflow/tensorflow/issues/999 nn.moments() - fix to handle axes that can't be automatically broadcast. - add a keep_dims optional argument to returned means and variances with the same dimensionality as the inputs. Change: 114056974
Vincent Vanhoucke
vanhoucke@google.com
1,454,857,380,000
59f1eba5fb94506a205fa2e81145667754739da5
Merge changes from GitHub, clean up linter errors, fix dependencies test. Change: 114064632
Martin Wicke
wicke@google.com
1,454,877,265,000
7bc6f528ec97a1153c5a12f4df6c2edf119d2f8c
Removing empty multiclass test as the case is tested by rnn test
Illia Polosukhin
ilblackdragon@gmail.com
1,454,884,132,000
597a6474649adcac487f0418cb6b985c155a71a2
Merge commit for internal changes
Martin Wicke
wicke@google.com
1,454,885,727,000
81cb3441df5acb428dbce25f5b287fb35d2076ff
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,291,000
bcefd8e0f553e384c79a439a2fcfe27624e9d223
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,445,000
3cd2b96eb96f392aea2c3684ca843404de59020d
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,489,000
f0ad2c179ecf8dbf1045b6ccd008d76a80867a67
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,520,000
b2852622ab9a555089ed7a7fa65c6064265d670d
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,601,000
aae44380256166127ece6f5010d4656556f5c60d
Remove duplicate import
Martin Wicke
martin.wicke@gmail.com
1,454,913,662,000
84150887f4234ec75b7819ff2b87b74622843266
Remove duplicate srcs_version
Martin Wicke
martin.wicke@gmail.com
1,454,913,711,000
effbc046c5fa28ca2a874f7defa7a6bcf4514c9a
Remove duplicate link
Martin Wicke
martin.wicke@gmail.com
1,454,913,777,000
b84e9efe63fe774e65bab27a3f657f54df26c333
Remove duplicate link
Martin Wicke
martin.wicke@gmail.com
1,454,913,806,000
a581a311d25b0b09cc684cf1875b1588755ffb4c
Remove duplicate link
Martin Wicke
martin.wicke@gmail.com
1,454,913,859,000
79f863dc3a132d84aec800755a823d78118d7af0
Merge pull request #1014 from tensorflow/cleanup-merge Fix merge gone wrong
Martin Wicke
martin.wicke@gmail.com
1,454,914,711,000
bddea00b76f1101ea392bd0bfb5254d8c2779395
Merge pull request #1010 from moutai/patch-1 fix minor filename typo
Martin Wicke
martin.wicke@gmail.com
1,454,914,757,000
30f179fc8a2967bceba7d8d0486bf8ed3d1829dc
Adding option for parallel build and serial tests In ci_parameterized_build.sh, allow the option of parallel bazel build followed by serial (--jobs=1) bazel test. Potentially useful for limiting the build time for GPU builds, where the tests cannot run in parallel due to contention or GPU memory.
Shanqing Cai
cais@google.com
1,454,816,170,000
44a646483d3157f86f5694209b5ebc727812cfad
Merge pull request #1011 from caisq/serial-tests-1.
Martin Wicke
wicke@google.com
1,454,914,925,000
108524e16776a8a529dad682dede57dbf68ae41e
Update README.md
Mandar
01.mandar@gmail.com
1,454,915,034,000
2e12c90afcc4b42d9eaccf09a5eca04d0d4625e8
Merge pull request #1003 from Mandar-Shinde/master.
Martin Wicke
wicke@google.com
1,454,915,056,000
1c54dde58197367c71319aa5a51f95d2a3faec6a
Add workspace root to the list of include path for swig compilation
Damien Martin-Guillerez
dmarting@google.com
1,454,701,016,000
3bf121cdda763970caca3eb0c2f5c4e13c15a67f
Merge pull request #1005 from damienmg/external.
Martin Wicke
wicke@google.com
1,454,915,241,000
e70489af6a40b3fdec3182564054bd9099a08711
Example code for SummaryIterator using `tf.train.summary_iterator`.
e3
eerwitt@gmail.com
1,454,794,208,000
0430539223cd9dd8108a13cbd62ceb6bfe1fa894
Merge pull request #1009 from eerwitt/example-summary-iterator-package.
Martin Wicke
wicke@google.com
1,454,949,824,000
11c7506bbb351cac785e139a7982fad49f3f1022
Fixed the fixed point code to make sure it compiles when AVX2 instructions are enabled
Benoit Steiner
benoit.steiner.goog@gmail.com
1,454,605,312,000
d18de4cbad8f04f5fe09961fac1f4f6eb8a4ba66
Merge pull request #991 from benoitsteiner/master.
Martin Wicke
wicke@google.com
1,454,975,221,000
c92af9cd4b9e44642a111394a65230252d62f4bf
Python3 install-test in pip.sh
Shanqing Cai
cais@google.com
1,454,956,150,000
a5739a852bf8efc796b1e497dc5c20c5fc56530f
Merge pull request #1022 from caisq/serial-tests-1.
Martin Wicke
wicke@google.com
1,454,981,114,000
7629ae60bf94fc0d39f40d0b7aeecff2f86a0ca0
Fix for python3 syntax incompatibility.
Povilas Liubauskas
liubauskas@gmail.com
1,454,959,088,000
9f722ecf06add30f6c0a3d560ad9811ef08435b9
Merge pull request #1023 from povi/master.
Martin Wicke
wicke@google.com
1,454,981,204,000
d797048c6e78d45b22bff31e96bfaaf36780087a
Remove tf.tensor_util and change tf.unsupported to tf.contrib.util tf.tensor_util was already not public (not documented), now it is not part of tf.__all__. The one function that was in tf.unsupported is now tf.contrib.util.constant_value, and there is also tf.contrib.util.make_tensor_proto. Change: 114118183
Geoffrey Irving
geoffreyi@google.com
1,454,952,849,000
77ac22f5226044f647d6185b77521b55023145eb
Implements c++ grad for StopGradient, Squeeze, Slice and Reverse. Change: 114120469
A. Unique TensorFlower
nobody@tensorflow.org
1,454,954,159,000
ff4e52b28dac0533d613973fe58a4b5b2888d349
Fix exception re-raising code. Exception instance e cannot be called to construct a new exception. Change: 114131158
A. Unique TensorFlower
nobody@tensorflow.org
1,454,959,544,000
3e33d444c6c0861d448755627982284736ed01a0
Generalize tf.image.random_crop to dimension-independent tf.random_crop The C++ 3-D-only RandomCrop op is now deprecated at GraphDef version 8, replaced with a python tf.random_crop that works for any dimension. This will allow random_crop to be used for other purposes. Unfortunately, tf.image.random_crop took 2 sizes rather than 3 for 3-D tensors. The new tf.random_crop always takes n sizes for rank n tensors; pass 3 as the last element if you want to not crop a last dimension of size 3. Change: 114135451
Geoffrey Irving
geoffreyi@google.com
1,454,961,764,000
7eae838548b2cc940b68c89d27e06fff46c78eb2
Changed the representation for pending and dead counts. Previously, each of these was a std::vector<int>*, sized to the number of nodes in the graph. This meant touching two cache lines every time we wanted to manipulate these values, and also required 8 bytes per node in total. Instead, we now use a PendingCounts helper class added in this change that uses just a single byte to represent both these counts in the common case of fewer than 7 outgoing edges. This helps with cache efficiency considerably because these values are heavily modified throughout the graph execution process by different threads running on different cores, and also because getting this to fit in, say, L1 cache instead of L2 cache, also helps efficiency on just a single core. As part of this, added a real interface that describes the various operations we want to do to these counts ("decrement_pending", "mark_live", "increment_dead_count", etc.), which is more understandable than having direct manipulation of counter values scattered throughout the ExecutorState code. Change: 114137346
A. Unique TensorFlower
nobody@tensorflow.org
1,454,962,788,000
6e86bf5ffc91b84f4df2c943d49942baaa3b21ba
Add tensorboard/lib/nanite, a simple system for storing data and creating Polymer data bindings in a typesafe fashion. Nanite exposes a single class, Store, which you can think of as a reified Polymer data binding. The store can bind out to a Polymer object (equivalent to [[downwardBinding]]) or bind up from a Polymer object (like {{oneWayAutomaticBinding}}). Unlike Polymer data bindings, there is a typesafe way to call these methods. Also, having the state accessibly stored in these objects will make it easier to serialize and deserialize the application state. Nanite also exposes some methods for transforming stores into other stores in a functional fashion. You can think of this as a very small system for FRP that is compatible with Typescript and Polymer. Nanite is thoroughly tested. N.B. I already have a version of the event dashboard that was refactored to use Nanite, so this definitely can work for our needs without too much effort :) Change: 114149110
A. Unique TensorFlower
nobody@tensorflow.org
1,454,969,963,000
f08592c1f8a26a2fad2c0c7ba309d4559c3c9369
Deleted an out of date comment. Change: 114149394
Benoit Steiner
benoit.steiner.goog@gmail.com
1,454,970,109,000
e5add48ffba4b32ef9939414d2b8bb30c2faaffd
Move requestManager out of components/tf-backend and into lib/js/requestManager. Change: 114154254
A. Unique TensorFlower
nobody@tensorflow.org
1,454,972,783,000
2585e0a75e2802ca8b9877fd06544ecca0b95cd9
TensorFlow C++ tests: Add a HostConstant to testlib. Change: 114156939
Eugene Brevdo
ebrevdo@gmail.com
1,454,974,229,000
5de788e467214c9056225680ae5075ea97be91ba
Typo in doc of unique_with_counts Change: 114158092
A. Unique TensorFlower
nobody@tensorflow.org
1,454,974,916,000
df66b9fe049cb17e58454f309b662bdaf0d14fdb
Reimplementation of TensorShape to significantly reduce its memory footprint for the common cases (not very many dimensions, dimension sizes fairly small). The old TensorShape implementation was 56 bytes, while this new implementation is 24 bytes in the common cases (and Tensor objects have a contained TensorShape object). The implementation has a 16 byte buffer, plus an int64 num_elements_ field. It switches between three representations that use the 16 byte buffer in different ways. The last two bytes of the 16 bytes are always a tag byte, indicating which representation is in use, and a number of dimensions byte that indicates the number of dimensions in the shape. This change also introduces a new restriction that tensor shapes can be at most 256 distinct dimensions. The different representations are: REP16: up to 7 dimensions stored inline, where each dimension value is less than 32768. REP32: up to 3 dimensions stored inline, where each dimension value is less than 2^31 - 1. REP_OUT_OF_LINE: the fallback case, where the inline bytes are used to store a pointer to a heap allocated gtl::InlinedVector<int64, 4> that holds the shape dimensions. Changed the signature of TensorShape::dim_sizes() to return an actual gtl::InlinedVector<int64, 4> rather than an ArraySlice<int64> (the old signature was problematic because it placed unnecessary requirements on the implementation, namely that it store the dimensions as a contiguous array of in64 values, which the new representation does not do for REP16 and REP32). Preserved the old implementation by renaming it TensorShapeOld and moving it to the tensor_shape_test.cc unittest, where it used for a randomized test to ensure that the behavior is the same. Most of the benefit is that the TensorShape object is smaller. Assignment is also generally faster, except for the rare REP_OUT_OF_LINE case, which is a bit slower (benchmark case /4): Run on machine with (40 X 2801 MHz CPUs); 2016/01/29-11:46:08 CPU: Intel Ivybridge with HyperThreading (20 cores) dL1:32KB dL2:256KB dL3:25MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------- BM_TensorShape_Assign/0 1 1 580669339 BM_TensorShape_Assign/1 1 1 588612006 BM_TensorShape_Assign/2 1 1 586977731 BM_TensorShape_Assign/3 1 1 564029390 BM_TensorShape_Assign/4 61 62 10000000 BM_TensorShapeOld_Assign/0 4 4 183910667 BM_TensorShapeOld_Assign/1 6 6 100000000 BM_TensorShapeOld_Assign/2 6 6 100000000 BM_TensorShapeOld_Assign/3 8 8 88330961 BM_TensorShapeOld_Assign/4 39 39 17701395 Change: 114162871
A. Unique TensorFlower
nobody@tensorflow.org
1,454,977,537,000
de64e5ae3e71946a5cb74722351097fee3bbf217
Rollback of camera permission request code (https://github.com/tensorflow/tensorflow/pull/926) that requires API level 23. Building the app for level 21 means the permission is granted automatically on install. Change: 114164638
A. Unique TensorFlower
nobody@tensorflow.org
1,454,978,544,000
172a805e604ce41f2e829c8792cb968803679f1e
Add google/protobuf:cc_wkt_protos as a dep of the cc_proto_library rule inside tf_proto_library. The python equivalent was already a dep of the py_proto_library. Change: 114165685
A. Unique TensorFlower
nobody@tensorflow.org
1,454,979,151,000
222feaa6ca3bbaf187232717ea73b08f4024fe52
Cleaning up pylints - removing some unnecessary ones, being more careful about re-enabling wildcard-import where appropriate. Change: 114167131
David G. Andersen
dga@google.com
1,454,980,004,000
92f29258496bc8acd49d653f8324a6090c178040
* test for other types of rnn
terrytangyuan
terrytangyuan@gmail.com
1,454,988,556,000
494f0fa77c1a5be7943871e0348b617ae69bf0a6
+ weights and biases in test to boost coverage
terrytangyuan
terrytangyuan@gmail.com
1,454,988,825,000
f0a7b869aedebc31d89b684b74ef537ea76cae94
+ comments for resnet
terrytangyuan
terrytangyuan@gmail.com
1,454,992,342,000
b80fb1a6bc0c6e7d43bd6c53a111b5eaa5e6593e
* table of contents for examples
terrytangyuan
terrytangyuan@gmail.com
1,454,993,408,000
90c86fce991fcd62eb46b9f9311b9604b26b7e08
- deleted accidentally included files
terrytangyuan
terrytangyuan@gmail.com
1,454,993,478,000
2197c067c79d1464eb957764f8edd120693dcdf6
Fix setup.py for Eigen headers Change: 114173506
Vijay Vasudevan
vrv@google.com
1,454,985,501,000
c977e835bd92ffa4f28ffd48f424a88cdc19db89
Adding a batch_size to predict method to split into smaller mini-batches if required
Illia Polosukhin
ilblackdragon@gmail.com
1,455,001,802,000
0c57c0391ed2d85c317dc5192d650aa8385fc254
Adding support to split numpy arrays into batches for prediction
Illia Polosukhin
ilblackdragon@gmail.com
1,455,002,891,000
d0ac8de3fc59db9e0b7cf32667965ed26f4d89b7
Updating resnet to save logdir, graph and checkpoints and run batches in prediction.
Illia Polosukhin
ilblackdragon@gmail.com
1,455,003,111,000
7cdfb0d3e9bc38da1804eda14c40b118ffd81f89
Added continue_training=True to the estimator in resnet to make sure it keeps training
Illia Polosukhin
ilblackdragon@gmail.com
1,455,003,298,000
9e35ac30735460136a2e997ef3408fc862f9c81f
added xrange to seq2seq_ops
Illia Polosukhin
ilblackdragon@gmail.com
1,455,004,034,000
a864d2497ef4926138d69a320cc3ac6a02f6e3a4
Adds GPU CUDA kernel for scatter ops.
Frank Chu
fpmc@google.com
1,453,937,039,000
55e87f198226cc84c9090b001fabbab187a4ef83
Merge branch 'master' into neural-translation
Illia Polosukhin
ilblackdragon@gmail.com
1,455,004,287,000
9bea801f178dc3a2fc47d796d27a746f0a6e1408
fix linters
Illia Polosukhin
ilblackdragon@gmail.com
1,455,004,482,000
d6ca2b7eae0b7f42d3e3f58451e4dd34f25d18f3
Merge branch 'resnet' into neural-translation
Illia Polosukhin
ilblackdragon@gmail.com
1,455,004,505,000
1627cc1fa04f0e41fadaf23f2daebe3b78bbaf72
Adding neural translation with words and language modeling on characters.
Illia Polosukhin
ilblackdragon@gmail.com
1,455,006,116,000
bb239fa22481e697dfb56c11d44d599e53360ac9
Using tf.nn.rnn* instead of direct import
Illia Polosukhin
ilblackdragon@gmail.com
1,455,006,343,000
d4b03ac200af381580b2fc115b9df4e0e6a67ffb
Merge pull request #782 from fpmchu/master.
Martin Wicke
wicke@google.com
1,455,006,424,000
e52003b01cd4e7e0bc7a698bcfec331f6d5638f1
Switch to enumerate in seq2seq ops
Illia Polosukhin
ilblackdragon@gmail.com
1,455,006,531,000
80cb6001ec9e217095846467d4855b370cf4db42
Add Docs blurb to howtos. also fixes redundant path elements in links.
Martin Wicke
martin.wicke@gmail.com
1,455,006,655,000
63cd67701e89b53bccac5ffb54d2ee5efe7e66fd
Merge pull request #1027 from tensorflow/docs-blurb.
Martin Wicke
wicke@google.com
1,455,006,699,000
5e879d6f29769b270c1d790a400be4d0fedf3a3f
More explicit directory structure instructions
Martin Wicke
martin.wicke@gmail.com
1,454,998,647,000
156dc452420a307d64ffd3d940ab182b3dbdf223
Merge pull request #1025 from tensorflow/contrib-instructions.
Martin Wicke
wicke@google.com
1,455,006,775,000
cc304733318f9635d37e72abc1c4d75b15a48d59
Add the Ops for the producer to MetaGraphDef. Change: 114175259
Josh Levenberg
josh11b@tensorflow.org
1,454,987,730,000
3d91f5585a46a8ac4de16b9f90fc0b1f703812e5
Don't cache kernels in op segment if the op is stateless. Hardens handling of ref types in identity removal logics. Change: 114178041
A. Unique TensorFlower
nobody@tensorflow.org
1,454,991,763,000
3b81ea93e314d0f7011dea4571e3caab89ab7038
Downgrade Android demo API level to 21 from 23, as no special 23 functionality is needed. This should also prevent camera permissions from needing to be explicitly requested. Change: 114189921
A. Unique TensorFlower
nobody@tensorflow.org
1,455,008,613,000
27c7a6bf7799a05c083700c7b7667542b3f1de5d
Fixing some test errors in Python3 PIP install The following tests fail in PIP test-on-install on Python3 and are fixed by this CL: tensor_array_ops_test.py py_func_test.py freeze_graph_test.py After this CL, there will be only one remaining test that fails in the Python3 PIP test-on-install, namely, saver_test.py Change: 114212677
A. Unique TensorFlower
nobody@tensorflow.org
1,455,028,825,000
d8ee3cbf769323a185614d9c427f809d135c9830
Make InTopK return False if any of the predictions are NaN or non-finite. Change: 114218051
A. Unique TensorFlower
nobody@tensorflow.org
1,455,033,082,000
fca72d03dea9192715713a8ee8e9877d45cc4264
s/max/std::max/ Change: 114218386
Manjunath Kudlur
keveman@gmail.com
1,455,033,336,000
75937824b9834b04c9a371447da27f322dfd9882
Merge commit for internal changes
Manjunath Kudlur
keveman@gmail.com
1,455,033,445,000
6ebacbcb18535bb22299ec3b5b3a5df227e12587
Added small test for bidirectional rnn and forked code from current github version of TF to make it work. Will cleanup when 0.7 is out
Illia Polosukhin
ilblackdragon@gmail.com
1,455,039,567,000
edc98e6269dc448f8e55be9af6018e4ee768b223
Merge branch 'master' of github.com:google/skflow
Illia Polosukhin
ilblackdragon@gmail.com
1,455,040,221,000
a131b40940717cadd45bc393faa75f6dc58ac53a
Merge branch 'resnet'
Illia Polosukhin
ilblackdragon@gmail.com
1,455,040,281,000
f4b0df86fc5120d260739f12af0b9723dfbbcf68
Improve the support for device functions when using `tf.import_graph_def()`. Previously the device function would run before the inputs to an op had been added. This yielded incorrect results for some device functions that depend on the inputs to an op. The fix is to split the creation of the op from applying its device functions (as we do for shape functions as well). Change: 114219160
Derek Murray
mrry@google.com
1,455,033,869,000
1bcebcc665e7e9c280be65899002fd3d5a7456a5
Fix a typo in tensor_shape.h Change: 114224625
Geoffrey Irving
geoffreyi@google.com
1,455,037,548,000
49e337b4f163aa0a570333da7933303096ba3d29
Make decode_jpeg automatically turn CMYK to RGB Our core/lib/jpeg now refuses to output 4 channel jpegs. It still accepts them as input, but unconditionally converts them to grayscale or RGB. Change: 114239759
Geoffrey Irving
geoffreyi@google.com
1,455,045,641,000
0c06d8d9aa88b0596b734c8feb6435d2cf359b7e
Merge commit for internal changes
Vijay Vasudevan
vrv@google.com
1,455,046,042,000
b0e7758cd25825b980cdff99faca7f888dab3d41
Merge branch 'master' into neural-translation
Illia Polosukhin
ilblackdragon@gmail.com
1,455,051,127,000
05454c61ea2290b99bde119640dd91fa56bb57de
Fix error in seq2seq ops after changing to enumerate
Illia Polosukhin
ilblackdragon@gmail.com
1,455,051,206,000
3c13ae058ea45d855d8029b3d19f6567b86430b5
TensorFlow: change int -> size_t to silence warning Change: 114243879
Vijay Vasudevan
vrv@google.com
1,455,047,764,000
27bbe92711a93613eca843772b6e7eb32ff96c35
Make the gfile package available when importing tensorflow. Update programs that were importing both 'tensorflow' and 'gfile' to use 'gfile' from the tensorflow import. Change: 114249943
A. Unique TensorFlower
nobody@tensorflow.org
1,455,051,406,000
8e34153636071b3f820b788d7f12513bad3da385
Removing need to import saver_pb2 inside restore function
Illia Polosukhin
ilblackdragon@gmail.com
1,455,060,648,000
45197e07c0dc410bde203462992841b457e0f61c
Add fully dynamic RNN op. Unofficial version (not yet documented). Current benchmarking results (speed, not memory): calculate forward and gradients w.r.t. variables Compiled -c opt --config=cuda CPU: Intel Ivybridge with HyperThreading (6 cores) dL1:32KB dL2:256KB dL3:12MB Tesla K40c Graph Creation: Static Unroll vs. Dynamic Unroll LSTM max_t dt(static) dt(dynamic) dt(dynamic)/dt(static) 1 0.713889 1.657085 2.321208 25 13.108961 1.784326 0.136115 50 25.887796 2.009615 0.077628 Calculation: Static Unroll with Dynamic Flow LSTM vs. Dynamic Unroll LSTM batch max_t units gpu dt(static) dt(dynamic) dt(dynamic)/dt(static) 256 50 512 False 2.098613 1.995831 0.951024 256 50 512 True 0.193702 0.218444 1.127730 256 50 256 False 0.630918 0.638142 1.011451 256 50 256 True 0.091755 0.119900 1.306731 256 50 128 False 0.278201 0.280812 1.009385 256 50 128 True 0.065415 0.093180 1.424446 256 100 512 False 4.252444 4.155129 0.977115 256 100 512 True 0.388840 0.459850 1.182622 256 100 256 False 1.247995 1.318395 1.056411 256 100 256 True 0.183765 0.238367 1.297130 256 100 128 False 0.558911 0.577376 1.033039 256 100 128 True 0.132579 0.192326 1.450656 512 50 512 False 3.682583 4.108835 1.115748 512 50 512 True 0.328870 0.378585 1.151168 512 50 256 False 1.129227 1.178769 1.043872 512 50 256 True 0.135209 0.163431 1.208724 512 50 128 False 0.435205 0.444609 1.021608 512 50 128 True 0.079627 0.106961 1.343268 512 100 512 False 7.183316 7.332909 1.020825 512 100 512 True 0.682864 0.722224 1.057641 512 100 256 False 2.033514 2.169560 1.066902 512 100 256 True 0.269665 0.338253 1.254347 512 100 128 False 0.776487 0.985389 1.269036 512 100 128 True 0.157085 0.207062 1.318155 Change: 114253188
Eugene Brevdo
ebrevdo@gmail.com
1,455,053,295,000
b27423b57a5a83992e3d066559abed4c555838f5
Make :mnist and :input_data public in examples/tutorials/mnist Change: 114254926
A. Unique TensorFlower
nobody@tensorflow.org
1,455,054,250,000
6dbf7f61b7f41e9bbfcd738dd73ade7fa498e8ca
Move private "framework_headers" target out of the public section of the BUILD file. Change: 114255368
Josh Levenberg
josh11b@tensorflow.org
1,455,054,496,000
d0a734f390342589c7b51a776a95e7673687be21
Added a utility method to annotate the GraphDef with inferred output shapes. Change: 114255626
Derek Murray
mrry@google.com
1,455,054,639,000
fb24124f40a7db61b3a7f7c7c23b69d4380242aa
TensorBoard should print descriptive error messages when it fails due to port or logdir issues, rather than just logging them. Change: 114256009
A. Unique TensorFlower
nobody@tensorflow.org
1,455,054,865,000
e52cecf69fa0efa782609a6d8f9494c08ae3aa7b
Added the use of google.protobuf.Any in meta_graph.proto to allow packing any any user protos. Change: 114256308
A. Unique TensorFlower
nobody@tensorflow.org
1,455,055,022,000