| Commit message (Collapse) | Author | Age |
|\
| |
| |
| | |
PiperOrigin-RevId: 209226085
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Also fix some simple issues that I saw when I benchmarked it (goes from ~3500 examples/sec -> ~4000 examples/sec)
- (nest) Expose is_mapping check that caches to python.
- (nest) Stop calling flatten when unnecessary in pack_sequence_as
- (nest) Set some functions to their swig wrappers directly (instead of wrapping them in another function)
- Directly call the gen_math_ops call in _aggregate_grads to skip any unnecessary python overhead.
- Stop falling back to slow path in _fast_fill.
PiperOrigin-RevId: 209223633
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209223543
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209222129
|
| |
| |
| |
| |
| |
| | |
I saw a MacOS flake after cl/208297005. Something platform-dependent seems to be going on.
PiperOrigin-RevId: 209220306
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209220232
|
| |
| |
| |
| |
| |
| | |
checkpoints of the model in user specified steps.
PiperOrigin-RevId: 209218489
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209217507
|
| |
| |
| |
| |
| |
| | |
Move more advanced notebooks into separate categories.
PiperOrigin-RevId: 209215629
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209212830
|
| |
| |
| |
| |
| |
| |
| |
| | |
Previously a Template's variables didn't show up in Model.weights, despite being available in Template.weights, Template.trainable_weights, and so on. Like other variable containers, these are now gathered (and forwarded through lists and dictionaries).
Also switches gathering to Model._layers from Model.layers (and likewise for CheckpointableDataStructures), since the un-filtered ._layers includes variables added directly to a container type (e.g. a Template). Updates and losses are still gathered from the post-processed/flattened .layers, which is fine for now since Layers are the only things that have updates or losses.
PiperOrigin-RevId: 209212023
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209209370
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209208985
|
| |
| |
| |
| |
| |
| | |
semantics)
PiperOrigin-RevId: 209208304
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209208090
|
| |
| |
| |
| |
| |
| |
| |
| | |
references.
This method is currently duplicated in Python code and will be used for TypeScript generation.
PiperOrigin-RevId: 209207139
|
| |
| |
| |
| | |
PiperOrigin-RevId: 209204056
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
* Redo functionalizing cond nodes by combining the cluster identification and flow forward parts.
* Propagate the predicates that must hold for a node to execute to handle cases with redundant switch/merge nodes.
- Predicates are now stored explicitly as an array (CondState);
* Each node has one CondId associated with its execution, this id represents a unique CondState;
* Introduce a merge and dead state to ensure correct behavior for one conditional feeding into another and for redundant switch nodes (i.e., switch on predicate p while already on a then/else branch of predicate p);
- The support for dead nodes is limited to case introduced by gradient code;
* Add Conditional class that represents a If that will be extracted;
* Extract branches reachable from Merge nodes;
- If the output of a switch does not feed into a merge, then that switch will be pruned - this is inline with expectation/while behavior;
- Extracting from the Merge nodes up can result in a switch node being shared between different Conditionals;
- Replicating constants also supported as workaround;
PiperOrigin-RevId: 209202945
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Please approve this CL. It will be submitted automatically, and its GitHub pull request will be marked as merged.
Imported from GitHub PR #21373
A new version of nsync is released, this updates the workspace.bzl and adds nsync to the libs that can be unbundled. I've packaged nsync on Gentoo Linux and tested the full unbundled build.
The CMake build is also updated to the latest version, the patch is no longer required and is removed. I have tested the cmake build on my machine but the CI needs to test windows since I can't.
The last commits update ./configure to handle adding TF_SYSTEM_LIBS and an option to disable stripping.
@m3bm3b, can you review too?
Copybara import of the project:
- 3d224b29fcf699b84cf538e2d16e3dd4e78e6a91 workspace: update nsync dep to 1.20.1 by Jason Zaman <jason@perfinion.com>
- 34536b88bedc5a83ad939438a49a3cc53aea210c CMake: nsync: update to 1.20.1 by Jason Zaman <jason@perfinion.com>
- f2444ba218ae280e03710f936b249b6b650ea3dd third_party: unbundle nsync by Jason Zaman <jason@perfinion.com>
- 4386d8c4b01020ba1c979258e5102ac1947d2574 configure: Add TF_SYSTEM_LIBS handling by Jason Zaman <jason@perfinion.com>
- f3fd1f110781cfb8cc941241a69eb41114787bfd Merge 4386d8c4b01020ba1c979258e5102ac1947d2574 into a05c7... by Jason Zaman <jasonzaman@gmail.com>
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/tensorflow/pull/21373 from perfinion:nsync 4386d8c4b01020ba1c979258e5102ac1947d2574
PiperOrigin-RevId: 209201545
|
|\ \
| | |
| | |
| | | |
PiperOrigin-RevId: 209198285
|
|\ \ \
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209198062
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209194778
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209187557
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209185706
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Fixes #19551
PiperOrigin-RevId: 209184588
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209180971
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
transformations.
PiperOrigin-RevId: 209174233
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Avoids whitespace changes in a later CL.
PiperOrigin-RevId: 209168698
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209168291
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
1. Remove mock from extra packages to install for testing since
it is already required to be installed for the build.
2. The instruction to build the build_pip_package script
is difficult to notice as it gets lost with
the Note in the GPU support section above it. As such,
add a subsection title for it.
RELNOTES: n/a
PiperOrigin-RevId: 209159059
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Makes sure there are read ops in the graph, to avoid errors on MetaGraph restore.
PiperOrigin-RevId: 209158129
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Additionally, assert that the same axis does not occur multiple times as Tensorflow Classic asserts.
PiperOrigin-RevId: 209132267
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Zero inputs for the multiples are supported but were being disallowed by an assertion.
PiperOrigin-RevId: 209132264
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
partially-specified shapes.
Fixes #21136
PiperOrigin-RevId: 209130870
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209119854
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This makes use of the newly added HLO support for grouped convolutions.
PiperOrigin-RevId: 209118249
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209105961
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
training ops as described in https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/quantize. This test help check for graphs that have training ops and raises a value error.
PiperOrigin-RevId: 209103265
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209099475
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209096783
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Previously, for equal comparisons the reported mismatched index would always be the last index in the shape rather than the index of the actual miscompare.
PiperOrigin-RevId: 209086635
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
checkpoints
A bit of extra infrastructure in checkpointable restore (save was already done) to support Python callbacks.
The same strategy should work for any Python state, although it's confined to non-pickled NumPy arrays at the moment.
PiperOrigin-RevId: 209085928
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209085003
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Add an overload which takes a random number generator to enable generation of different random values in sequential calls to MakeFakeArguments. Add a mechanism for generating arrays of unique values (no duplicates) for the key inputs to key/value kSorts. Remove some sorcery in generating float arrays and replace with a uniform distribution. The underlying reason for using this strange distribution no longer exist.
PiperOrigin-RevId: 209083904
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209082119
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
ParameterServerStrategy
PiperOrigin-RevId: 209078920
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This is useful when debugging auto-jit issues.
PiperOrigin-RevId: 209078538
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209077767
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209075749
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 209074011
|