| Commit message (Collapse) | Author | Age |
... | |
|
|
|
|
|
| |
occasionally be NaN.
Change: 142669425
|
|
|
|
|
|
|
|
|
| |
* if user defined an init_op which should not run after restoring. Current implementation caused to run init_op since session_manager thinks nothing is restored.
Removed variables_to_restore argument since it's usage is conflicting with Scaffold.saver. If user defined a saver and variables_to_restore, current implementation silently ignores variables_to_restore.
Used specific checkpoint file feature of session-manager which is recently added.
Change: 142669254
|
|
|
|
|
|
| |
it is no longer used.
Change: 142668381
|
|
|
|
|
|
| |
tensors as learning parameters in addition to just floats.
Change: 142664885
|
|
|
|
|
| |
Always use `OP_REQUIRES_OK_ASYNC(..., done)` in an AsyncOpKernel instead of `OP_REQUIRES_OK()`, or else the `done` callback will not be called in the event of a non-OK status.
Change: 142664282
|
|
|
|
|
|
| |
partial tensor shapes for exact equality.
Change: 142660788
|
|
|
|
|
| |
the event manager.
Change: 142660218
|
|
|
|
| |
Change: 142647790
|
|
|
|
| |
Change: 142647250
|
|
|
|
|
|
| |
- compute_boundary_ts was sometimes adding spurious inputs
- sgv.consumers was returning op inside the subgraph
Change: 142645987
|
|
|
|
| |
Change: 142643418
|
|
|
|
| |
Change: 142623343
|
|
|
|
|
| |
free UI up for interaction)
Change: 142623328
|
|
|
|
|
|
| |
itself in tensorflow/cc/framework/scope.cc
Change: 142615345
|
|
|
|
|
|
| |
latest gemmlowp, which makes it an error by default to build without
SSE4.1 on x86 or NEON on ARM.
Change: 142614852
|
|
|
|
| |
Change: 142614569
|
|
|
|
|
|
| |
updates are not yet supported. There still might be some bugs in the libxsmm
code for backward convolutions, so that mode is not recommended for use yet. Since this feature is experimental, "--define tensorflow_xsmm_backward=1" needs to be used on the build command line to enable it.
Change: 142614258
|
|
|
|
| |
Change: 142613993
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The `WorkerCacheInterface` class recently gained a `ReleaseWorker()`
that is designed to allow the (now misnamed) `CreateWorker()` to
return a shared pointer. This feature is used in
`rpc_rendezvous_mgr.cc` to cache WorkerInterface instances, but
elsewhere the raw destructor is used, which makes it unsafe to
implement a caching WorkerCacheInterface elsewhere. Since this would
be useful for implementing local worker cut-through, this CL enforces
the consistent use of `ReleaseWorker()` for all `WorkerInterface`
pointers.
Change: 142611501
|
|
|
|
| |
Change: 142610795
|
|
|
|
|
| |
Fix incorrect spelling of 'euclidean'.
Change: 142607588
|
|
|
|
|
|
|
|
|
|
|
| |
This is the first pass at adding mouse support to tfdbg CLI. It focuses on the analyzer part of the CLI. All these changes are backward compatible, in the sense that all keyboard commands continue to work as before.
* The list_tensors (lt) command now has hyperlink-like clickable text in its output, which users can use to navigate to the tensor value display with only one click.
* The outputs from all analyzer commands (lt, pt, ni, li, lo) have a main menu displayed near the top, right below the title bar. The menu contain items such as list_tensors, node_info, print_tensor, which get enabled and disabled properly depending on the context.
Also in this CL:
* Let users print tensors using the node name (instead of the tensor name), if the node has generated only one dump from one output slot.
Change: 142605017
|
|
|
|
| |
Change: 142601309
|
|
|
|
| |
Change: 142600828
|
|
|
|
|
|
| |
from the Python API.
Change: 142598006
|
|
|
|
| |
Change: 142595367
|
|
|
|
|
|
|
|
|
|
|
|
| |
The bug has to do with how node vs. tensor names are handled by the pending input tracker. The issue was as follows: suppose a node called A produces two outputs on two slots, A:0 and A:1. B receives inputs from A:0 and C receives input from A:1.
Suppose only the two following tensors are watched by the debugger: {A:0, C:0}, then there is no guarantee that the debug tensor for A:0 is dumped before the debug tensor of C:0, because C doesn't receive input from A:0 (it receives input from A:1!). As a result, during causality validation, the debugger may sometimes see that C:0 is dumped before A:0, thinking its a causality violation.
This CL fixes the bug by tracking the inputs down to the level of output slots, instead of just at the level of nodes.
Hitchhiking changes:
* Minor change in title style of the offline debugger.
Change: 142593110
|
|
|
|
| |
Change: 142591245
|
|
|
|
|
|
|
|
|
|
|
| |
This enables each partition GraphDef to be swapped into the
RegisterGraphRequest without making a copy.
The StatsPublisherInterface implementation is now responsible for
taking a copy of the GraphDef if it wants to store it (e.g. for later
visualization); by default, no copy is retained, since there is no
mechanism for accessing it.
Change: 142589346
|
|
|
|
| |
Change: 142588049
|
|
|
|
| |
Change: 142582442
|
|
|
|
| |
Change: 142578639
|
|
|
|
| |
Change: 142577517
|
|
|
|
|
|
| |
the tests for finer granularity. Allow the read queue to be automatically sized in a way that provides better guarantees that the input will truly be shuffled.
Change: 142576906
|
|
|
|
| |
Change: 142575145
|
|
|
|
|
| |
with no nodes (empty function bodies).
Change: 142573865
|
|
|
|
| |
Change: 142571460
|
|
|
|
| |
Change: 142567058
|
|
|
|
| |
Change: 142565549
|
|
|
|
| |
Change: 142557663
|
|
|
|
| |
Change: 142517940
|
|
|
|
| |
Change: 142517678
|
|
|
|
|
|
|
|
|
| |
less accurate one. This enables, but does not yet provide, significant speed improvements for image models bottlenecked on image-preprocessing.
RELNOTES: tf.image.decode_jpeg by default uses the faster DCT method, sacrificing
a little fidelity for improved speed. One can revert to the old
behavior by specifying the attribute dct_method='INTEGER_ACCURATE'.
Change: 142517485
|
|
|
|
| |
Change: 142515779
|
|
|
|
| |
Change: 142508618
|
|
|
|
| |
Change: 142507111
|
|
|
|
| |
Change: 142501037
|
|
|
|
| |
Change: 142500394
|
|
|
|
| |
Change: 142500365
|
|
|
|
| |
Change: 142495971
|