- 10 1月, 2017 1 次提交
-
-
由 gunan 提交于
* Update release notes for 1.0.0-alpha. * Add numpy compatibility to release notes.
-
- 08 1月, 2017 5 次提交
-
-
由 gunan 提交于
Setting losses_test to medium since it times out on Jenkins at times.
-
由 A. Unique TensorFlower 提交于
Change: 143740377
-
由 Gunhan Gulsoy 提交于
-
由 Gunhan Gulsoy 提交于
-
由 Gunhan Gulsoy 提交于
-
- 06 1月, 2017 3 次提交
-
-
由 gunan 提交于
Merge master into r1.0 to fetch all the fixes.
-
由 Gunhan Gulsoy 提交于
-
由 Rohan Jain 提交于
Branch 143639671
-
- 05 1月, 2017 31 次提交
-
-
由 Shanqing Cai 提交于
-
由 Martin Wicke 提交于
Change: 143639671
-
由 Andrew Hundt 提交于
* worspace.bzl uses zlib permalink zlib permalinks including the current version are always in http://zlib.net/fossils/zlib-*.*.*.tar.gz links in http://zlib.net/zlib-*.*.*.tar.gz are removed when no longer current. Should resolve #6594 if merged into r0.12 branch. * Remove comment, moving to git commit description.
-
由 A. Unique TensorFlower 提交于
Change: 143630705
-
由 Martin Wicke 提交于
Change arg order for {softmax,sparse_softmax,sigmoid}_cross_entropy_with_logits to be (labels, predictions), and force use of named args to avoid accidents. Change: 143629623
-
由 A. Unique TensorFlower 提交于
Change: 143626185
-
由 Illia Polosukhin 提交于
Update tf.contrib.framework documentation to contain checkpoint utils. Change: 143624505
-
由 Derek Murray 提交于
This CL includes wrapper classes for the protocol buffer messages `tensorflow::RunStepRequest` and `tensorflow::RunGraphRequest`. Previously the service arguments were always protocol buffer messages, which can entail copying large tensor values into and out of the request message. This change makes the backend code deal with abstract `tensorflow::RunStepRequestWrapper` and `tensorflow::RunGraphRequestWrapper` interfaces and adds three concrete implementations of each interface: * An mutable in-memory wrapper, which maintains the tensor data in `tensorflow::Tensor` objects, and provides the most efficient implementation when the client and master are in the same address space. * A mutable protobuf wrapper, which has a similar implementation to today's client code. * A const wrapper around a const protobuf, which has a similar implementation to today's server code. This is another improvement for issue #6256. Change: 143620823
-
由 Nathan Silberman 提交于
Change: 143619790
-
由 Suharsh Sivakumar 提交于
This is only needed for PartialRun and slows down the Run path. Change: 143618516
-
由 Benoit Steiner 提交于
Change: 143617892
-
由 Li Lao 提交于
Change: 143612122
-
由 A. Unique TensorFlower 提交于
Defer optimizer function run in linear classifier until apply gradient step. This keeps the worker to get correct global step through contrib_variables.get_global_step. This defer also enables the use of a customized learning rate in optimizer, e.g. exp decay learning rate in linear classifier. Otherwise it will cause "global_step is required" error when runs with replicas. Change: 143602114
-
由 A. Unique TensorFlower 提交于
Removing comments for investigation of the root cause of test tolerance change. Turns out it is expected: originally a more strict tolerance was used incorrectly where it has been known to be too restrictive for other similar test cases when expanding the tests to cover the added projection operation. Change: 143601145
-
由 Vijay Vasudevan 提交于
Fixes #6602. Change: 143600453
-
由 A. Unique TensorFlower 提交于
getting deprecated. The call signatures are changed from: tf.concat(concat_dim, value, name) to tf.concat_v2(value, axis, name). Change: 143596582
-
由 gunan 提交于
Remove unused TF_NEED_SYCL from ./configure.
-
由 A. Unique TensorFlower 提交于
Change: 143589458
-
由 A. Unique TensorFlower 提交于
Change: 143589065
-
由 Patrick Nguyen 提交于
Change: 143587587
-
由 Peter Hawkins 提交于
Change: 143586818
-
由 A. Unique TensorFlower 提交于
Change: 143585660
-
由 gunan 提交于
Break indirect dependency on contrib for windows.
-
由 Andrew Harp 提交于
Change: 143583137
-
由 Ian Langmore 提交于
This operator is initialized with a num_rows and batch_shape parameter. An alternative would be a single shape parameter. The reasons for this are: 1. Compatibility with tf.eye and np.eye 2. If "shape" was used, and ndims was not known statically, then operator.apply(x) and operator.solve(x) need to force a broadcast by adding an array of zeros. This is unnecessarily inefficient. Having an explicit "batch_shape=None" option allows (in the event that batch_shape is None) operator.apply(x) and operator.solve(x) to simply return x. Change: 143582824
-
由 A. Unique TensorFlower 提交于
Change: 143580493
-
由 Joshua V. Dillon 提交于
Previously it was assumed overrides would be statically on or off (either way the override itself could have been non-static). This cl allows for fully dynamic specification of batch_shape and/or event_shape override, ie, we no longer require that the decision to override shape be known during graph construction. Change: 143576091
-
由 Andrew Harp 提交于
Android: add debug-specific overlay for detection activity that can be rendered on a different schedule from tracking overlay, improving performance Change: 143575879
-
由 A. Unique TensorFlower 提交于
Change: 143574390
-
由 A. Unique TensorFlower 提交于
Change: 143570831
-
由 Pete Warden 提交于
Change: 143570811
-