| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
from hlo_constant_folding to hlo_evaluator.
Additionally:
- In ShapeUtil::ForEachIndex:
* fix a bug where visitor is called when the shape has zero elements (e.g., F32{1,0})
* added test case for ForEachIndex.
- In HloEvaluator:
* Instead of copying and caching a Constant instruction, return the literal directly if the instruction is constant.
* Fix an issue where TUPLE and OPAQUE primitives are not keyed in the templated typed_visitor.
* Use (fixed) LiteralUtil::Populate to populate resulting literal, fixes the preexisting bug in the evaluator where R0 and shape with zero size dimensions are not handled.
* Refactor ElementWiseUnaryOp and HandleCompare to be templatized on the operand's type.
* Refactor IsFinite to be top level since it is only applicable to floats and the return type is always boolean.
* Change from std::remainder to std::fmod for kRemainder to be compliant with existing XLA behavior.
* Change from std::max and std::min to std::fmax and std::fmin to handle NaNs.
* Minor comments fix.
- Disables constant_folding and reshape-motion for ClientLibraryTestBase so that constant folding would not affect the intended code paths to be execercised by the test. In the longer term we plan change all Constants to Parameter and re-enable constant_folding in tests.
PiperOrigin-RevId: 157174708
|
|
|
|
| |
PiperOrigin-RevId: 157169680
|
|
|
|
| |
PiperOrigin-RevId: 157169178
|
|
|
|
|
|
|
|
| |
shape and type information for multiple tensors.
Apply this to QueueDequeueV2 handled by grappler.
PiperOrigin-RevId: 157163757
|
|
|
|
|
|
|
|
|
| |
them to work with inputs of undefined rank.
Add test for `indicators_to_sparse_ids` `dtype` arg.
Small update to `unstack` pydoc.
PiperOrigin-RevId: 157160634
|
|
|
|
|
|
| |
loading/remapping embeddings, warm-starting weights/biases, etc.)
PiperOrigin-RevId: 157148893
|
|
|
|
| |
PiperOrigin-RevId: 157147578
|
|
|
|
| |
PiperOrigin-RevId: 157146520
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This pattern tests all tests twice (once for use_gpu=True, once for
use_gpu=False).
Switch to using a single test pattern that uses use_gpu=True,
which tests on GPU if a GPU is available, and tests on CPU otherwise.
the %test%_gpu target tests on GPU, the %test% target tests on CPU, so
both are covered.
The tricky situation is when you want to test locally on GPU, you have
to remember to run the _gpu target, but reducing test time is more
important than individual developers making mistakes at this point.
PiperOrigin-RevId: 157141876
|
|
|
|
| |
PiperOrigin-RevId: 157129614
|
|
|
|
|
|
| |
Subsequent changes will make TensorBoard use this images plugin instead of the previous handlers for image-related data.
PiperOrigin-RevId: 157123389
|
|
|
|
|
|
| |
shape of a scalar when the shape hasn't been set.
PiperOrigin-RevId: 157120482
|
|
|
|
|
|
| |
The deleted comment suggests that CUDA 6.5 is enough here although we're now building binaries with CUDA 8. The updated error message makes it obvious when CuDNN fails due to an outdated driver.
PiperOrigin-RevId: 157118560
|
|
|
|
|
|
| |
Given" instead of "_FeatureColumn.Given".
PiperOrigin-RevId: 157116424
|
|
|
|
|
|
|
|
| |
of their names
This is a step towards switching to the new session API, which expects the corresponding C objects instead of names.
PiperOrigin-RevId: 157114094
|
|
|
|
| |
PiperOrigin-RevId: 157105234
|
|
|
|
|
|
| |
evaluation metrics computed along with the best validation metric.
PiperOrigin-RevId: 157081910
|
|
|
|
| |
PiperOrigin-RevId: 157070223
|
|
|
|
| |
PiperOrigin-RevId: 157068219
|
|
|
|
|
|
|
|
|
|
| |
couple of reasons:
1) Shapes are useful for debugging whether the grappler operations match the expected operations. For example, checking that the batch sizes propagate to the individual matrix multiply.
2) The counts are useful for matching the number of individual grappler ops against the actual op counts from a timeline profile. This will help in debugging the cost that grappler outputs.
3) The individual per-op timings give a more detailed breakdown of how grappler is estimating the run time.
PiperOrigin-RevId: 157062681
|
|
|
|
| |
PiperOrigin-RevId: 157062094
|
|
|
|
|
|
| |
extra checks clutter up code.
PiperOrigin-RevId: 157061530
|
|
|
|
|
|
|
|
| |
inside a control-flow scope, such as tf.cond() or tf.while_loop().
Fixes #8604.
PiperOrigin-RevId: 157060591
|
|
|
|
|
|
|
|
|
|
|
| |
- change OperandCanTriviallyChangeShape to only look at operand's opcode, instead of comparing the dimension of the shape between the instruction and its operand.
- remove accounting for non-trivial reshapes/transposes, and instead move the requirement to FirstNonScalarReshapeOperand instead.
This fixes an issue where a elementwise op's reshape0(constant0) operand gets moved (because reshape0 and constant0 are of different shapes), even when there is no benefit is doing so.
Also fixes incorrect test comments.
PiperOrigin-RevId: 157060319
|
|
|
|
|
|
|
|
|
|
|
| |
This removes the circular dependency when creating a HloModule, then a
HloModuleConfig with the help of the module's entry computation, then assigning
the config back to the module. Now we have to pass a config when creating a
module, or a default config gets created.
This allows removing quite a bit of boilerplate code.
PiperOrigin-RevId: 157059949
|
|
|
|
|
|
|
| |
** GPU transfer manager and GPU specific infeed manager/infeed buffer implementation
** Infeed thunk
PiperOrigin-RevId: 157054373
|
|
|
|
| |
PiperOrigin-RevId: 157052586
|
|
|
|
| |
PiperOrigin-RevId: 157049830
|
|
|
|
| |
PiperOrigin-RevId: 157047588
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
tfprof> op -select micros,bytes,occurrence -order_by micros
SoftmaxCrossEntropyWithLogits 36.58MB (100.00%, 0.05%), 1.37sec
(100.00%, 23.56%), 30
MatMul 2720.57MB (99.95%, 3.66%), 988.90ms (76.44%,
17.05%), 3450
ConcatV2 741.37MB (96.29%, 1.00%), 421.44ms
(59.38%, 7.27%), 6098
Mul 3957.24MB (95.29%, 5.33%), 418.90ms
(52.12%, 7.22%), 9427
Add 740.05MB (89.96%, 1.00%), 335.26ms
(44.89%, 5.78%), 2180
Sub 32.46MB (88.97%, 0.04%), 216.44ms
The CL also:
1. Unify code view and op view implementation and rename.
2. Clean up a few unused features.
3. Unify the option semantics of all views.
PiperOrigin-RevId: 157043043
|
|
|
|
|
|
| |
help parsing for classifiers.
PiperOrigin-RevId: 157041528
|
|
|
|
| |
PiperOrigin-RevId: 157040972
|
|
|
|
| |
PiperOrigin-RevId: 157039326
|
|
|
|
| |
PiperOrigin-RevId: 157038372
|
|
|
|
|
|
|
|
| |
This was originally added a long time ago to help with GPU benchmarks; I think
it adds needless complexity to the Backend class, which we intend to simplify.
If this is still an issue for benchmarks, it should be done on a higher level.
PiperOrigin-RevId: 157037991
|
|
|
|
| |
PiperOrigin-RevId: 157037658
|
|
|
|
| |
PiperOrigin-RevId: 157037270
|
|
|
|
|
|
|
| |
Fixes #4330.
RELNOTES: Allow uses of over-parameterized separable convolution.
PiperOrigin-RevId: 157035904
|
|
|
|
| |
PiperOrigin-RevId: 157035658
|
|
|
|
| |
PiperOrigin-RevId: 157034590
|
|
|
|
|
|
|
| |
debug_utils is an internal target that shouldn't be used by users.
Replacing it with the correct debug / tf_debug import.
PiperOrigin-RevId: 157032842
|
|
|
|
| |
PiperOrigin-RevId: 157032759
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This CL achieves the following:
- The output of the /data/runs endpoint no longer contains any
references to "scalars."
- The router no longer has any special handling for scalars.
- The application (application.py) and its tests no longer have any
special handling for scalars.
As part of the process of pulling out the scalars dashboard, this leaves
behind some assorted infrastructure to make it easier to migrate other
plugins (primarily, in the router).
Tests done so far:
- All tests descended from the "tensorboard" directory
- Manual interaction of the following forms: expanding a scalar plot;
enlarging a scalar plot; shrinking a scalar plot; using the run
selector's check boxes and radio buttons; using the run selector's
regex filter; creating a tag group; viewing and scrolling through
images, along with a casual check of the other dashboards
- Manual inspection of the /data/runs output to make sure that it
contains no scalar data
PiperOrigin-RevId: 157028690
|
|
|
|
| |
PiperOrigin-RevId: 157020704
|
|
|
|
|
|
|
|
| |
This is a fully backwards compatible change that allows the user to decide:
* Which axes to normalize
* The shape of the scale/center variables (these are broadcasted).
PiperOrigin-RevId: 157019098
|
|
|
|
| |
PiperOrigin-RevId: 157018996
|
|
|
|
| |
PiperOrigin-RevId: 157016549
|
|
|
|
|
|
|
|
| |
clear that it is a list of aliases.
Add parent class to class pages.
PiperOrigin-RevId: 157011127
|
|
|
|
| |
PiperOrigin-RevId: 157011106
|
|
|
|
| |
PiperOrigin-RevId: 157010139
|