Commits

Branch selector

 master

User selector

rxsang

Datepicker

All time

Commit History

Commits on Oct 19, 2019

• Also propagate `is_mirrored` annotation for ONLY_FIRST_REPLICA TPUMirroredVariable.

  Show description for 07c8634
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  07c8634

  Copy full SHA for 07c8634

Commits on Oct 18, 2019

• Avoid re-evaluating batch_size in BatchDataset. This will help the startup time performance in remote eager, as each evaluation is underlying calling EagerTensor.numpy(), which is a blocking point.

  Show description for 5f47cee
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  5f47cee

  Copy full SHA for 5f47cee

Commits on Oct 15, 2019

• Fix TPUStrategy experimental_local_results may cause outputs in disorder, if the alphabetical order and the actual order of devices are different.

  Show description for 9d7bd16
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  9d7bd16

  Copy full SHA for 9d7bd16

Commits on Oct 11, 2019

• Allow calling tf.config.experimental_connect_to_cluster multiple times.

  Show description for 1db559b
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  1db559b

  Copy full SHA for 1db559b

Commits on Oct 10, 2019

• Only propagate "is_mirrored_variable" tag for mirrored variables which values are guaranteed to be the same all the time.

  Show description for 090c309
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  090c309

  Copy full SHA for 090c309

Commits on Oct 4, 2019

• Refine the client-server version mismatch error message.

  Show description for f7ffed0
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  f7ffed0

  Copy full SHA for f7ffed0

Commits on Oct 3, 2019

• Improves function tracing time of variables created inside functions. In _initialize_uninitialized_variables, pack all variable_is_initialized ops into one tensor and convert it to numpy in one cal…

  Show description for 66f1ac5
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  66f1ac5

  Copy full SHA for 66f1ac5

Commits on Oct 2, 2019

• Add an escape path for people to skip dynamic padding in TPUStrategy.

  Show description for 81f2b3d
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  81f2b3d

  Copy full SHA for 81f2b3d

Commits on Sep 24, 2019

• Place padding ops on TPUs if the original input nodes has already been on TPUs, so XLA op demand mode can be triggered. This will avoid unnecessary data copy between host and device.

  Show description for 2405e8a
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  2405e8a

  Copy full SHA for 2405e8a

Commits on Sep 4, 2019

• Register XLA device kernel for _SwitchN op.

  Show description for f33b2e0
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  f33b2e0

  Copy full SHA for f33b2e0

Commits on Aug 27, 2019

• Add "is_mirrored_variable" attribute into TPUReplicatedInput op.

  Show description for f1122b0
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  f1122b0

  Copy full SHA for f1122b0

Commits on Aug 24, 2019

• Only truncate ".brain" suffix in eager mode.

  Show description for dac7335
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  dac7335

  Copy full SHA for dac7335

Commits on Aug 22, 2019

• Add an `enter_master_device` flag in tf.config.experimental_connect_to_cluster API.

  Show description for 2800f9b
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  2800f9b

  Copy full SHA for 2800f9b

• Support running computation on a subset of cores in TPUStrategy.

  Show description for ca82684
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  ca82684

  Copy full SHA for ca82684

Commits on Aug 16, 2019

• Make sure TPUStrategy handles inputs with static shapes but different sizes correctly. Fix the issue padding_map is not populated in this case.

  Show description for fe77010
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  fe77010

  Copy full SHA for fe77010

Commits on Aug 14, 2019

• A workaround to make sure strategy.reduce work correctly with dynamic shapes.

  Show description for 4183b69
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  4183b69

  Copy full SHA for 4183b69

Commits on Aug 9, 2019

• Add load and save model tests for TPU.

  Show description for 5536146
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  5536146

  Copy full SHA for 5536146

Commits on Aug 3, 2019

• Automated rollback of commit 41884bf3d3c29f1520ba84a5233e30cb863f736e

  Show description for 37ef75e
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  37ef75e

  Copy full SHA for 37ef75e

Commits on Aug 2, 2019

• Enter the tpu worker device scope automatically if within TPUStrategy scope. This avoids the case that user always need to manually place their dataset and functions on tpu worker.

  Show description for 41884bf
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  41884bf

  Copy full SHA for 41884bf

Commits on Jul 23, 2019

• Fix a graph lifting bug which the same init op may be copied multiple times during the lifting. This may cause TPUMirroredVariable on different devices initialized to different values.

  Show description for 745b24b
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  745b24b

  Copy full SHA for 745b24b

Commits on Jul 18, 2019

• Handle zero batch case in BatchNorm only when 'experimental_enable_get_next_as_optional' is set in strategy object.

  Show description for 1d5b243
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  1d5b243

  Copy full SHA for 1d5b243

Commits on Jul 17, 2019

• Update all the TPUStrategy examples to use the new v2 APIs, i.e.

  Show description for eba2610
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  eba2610

  Copy full SHA for eba2610

Commits on Jul 15, 2019

• Re-enable keras lstm tests.

  Show description for 927ddfd
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  927ddfd

  Copy full SHA for 927ddfd

• Relax tpu.replicate to handle inputs with static shapes but different among replicas. This is useful in TF2.0, when user calls iterator.get_next() outside tf.function, tpu.replicate will receive ea…

  Show description for e926ed2
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  e926ed2

  Copy full SHA for e926ed2

Commits on Jul 11, 2019

• Only enable last partial support in MirroredStrategy in eager mode.

  Show description for 2cb25a8
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  2cb25a8

  Copy full SHA for 2cb25a8

Commits on Jun 28, 2019

• Only enable optional_get_next when the iterator output_shapes are not static.

  Show description for 750b4f3
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  750b4f3

  Copy full SHA for 750b4f3

• Set context.MIRRORING_ALL by default in TPUStrategy, this is to make sure TPU Pod gets good performance in TF2.0.

  Show description for 645b3cb
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  645b3cb

  Copy full SHA for 645b3cb

Commits on Jun 27, 2019

• Add subgraph device into the dumped partitioned graph name for easier debugging.

  Show description for 300590a
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  300590a

  Copy full SHA for 300590a

Commits on Jun 20, 2019

• Add more TraceMe to track function runtime overhead.

  Show description for c2890a0
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  c2890a0

  Copy full SHA for c2890a0

Commits on Jun 13, 2019

• Comment fix for tpu symbols.

  Show description for 1ae6a2a
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  1ae6a2a

  Copy full SHA for 1ae6a2a

Commits on Jun 12, 2019

• Use device_assignment in TPUStrategy.

  Show description for 98c6137
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  98c6137

  Copy full SHA for 98c6137

• Fix a bug in TPUMirroredVariable that variable handle ids may collide if variables are created inside tf.function but outside is eager mode.

  Show description for 5228dca
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  5228dca

  Copy full SHA for 5228dca

Commits on Jun 11, 2019

• Re-enable iterating over DistributedDataset inside tf.function TPU test.

  Show description for c148a2d
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  c148a2d

  Copy full SHA for c148a2d

Commits on Jun 8, 2019

• Fix the issue that customized Sonnet module doesn't work on TF2.0 TPU. The following changes are included:

  Show description for 63b9442
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  63b9442

  Copy full SHA for 63b9442

Commits on May 30, 2019

• Fix TPU initialization in eager mode.

  Show description for 32beaef
  

  rxsang

  authored and

  tensorflower-gardener

  committed
  32beaef

  Copy full SHA for 32beaef

Pagination

Previous

Next