| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
GPU would fail due to having too many parameters to fit in memory because Concat's signature is variadic and can have an unlimited number of inputs.
PiperOrigin-RevId: 211942734
|
|
|
|
| |
PiperOrigin-RevId: 211942571
|
|
|
|
|
|
|
|
|
| |
I put this in as a quick hack because init_value is usually a constant, but
it's really easy to construct a case where it's not. The code also became more
complex because of the constant buffer work, sharing that with the fused IR
emitter is a good thing.
PiperOrigin-RevId: 211936337
|
|
|
|
|
|
|
|
| |
with default changed to SUM_OVER_BATCH_SIZE
This would involve making changes to all existing uses of RNNClassifier to set the loss reduction argument explicitly to SUM (previous default was SUM)
PiperOrigin-RevId: 211917502
|
|
|
|
|
|
|
|
| |
I don't believe there is currently a use-case for a different VSpace (and it doesn't seem to be controllable through any public method).
If it is a usecase we want to support, it should be simple enough to add an overload of TFE_Py_TapeGradient.
PiperOrigin-RevId: 211917235
|
|
|
|
| |
PiperOrigin-RevId: 211916428
|
|
|
|
|
|
| |
epsilon controls the maximum number of quantiles generated.
PiperOrigin-RevId: 211914388
|
|
|
|
| |
PiperOrigin-RevId: 211912785
|
|
|
|
|
|
|
| |
While there, tweak the implementation of convolution in the HLO evaluator to be
a little simpler.
PiperOrigin-RevId: 211911253
|
|
|
|
| |
PiperOrigin-RevId: 211907050
|
|
|
|
| |
PiperOrigin-RevId: 211906579
|
|
|
|
|
|
| |
cuDNN 7.1.4 and 7.2 has non-determinisic bug if the buffer is not zeroed.
PiperOrigin-RevId: 211905127
|
|
|
|
|
|
| |
optimizers and setting optimizer names.
PiperOrigin-RevId: 211900252
|
|
|
|
| |
PiperOrigin-RevId: 211899762
|
|\
| |
| |
| | |
PiperOrigin-RevId: 211897876
|
| |
| |
| |
| |
| |
| |
| |
| | |
Model.fit() function to support taking in dictionaries of features in.
Support for functional models coming in a subsequent change.
PiperOrigin-RevId: 211897153
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211896300
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211895566
|
| |
| |
| |
| |
| |
| | |
Plumb these through to RecordWriterOptions
PiperOrigin-RevId: 211894734
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211892456
|
| |
| |
| |
| |
| |
| | |
based runtime in Swift for Tensorflow.
PiperOrigin-RevId: 211892308
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211892283
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211891325
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211890783
|
| |
| |
| |
| |
| |
| |
| |
| | |
`tf.data.Dataset.interleave`.
Unlike the `tf.data.contrib.parallel_interleave` whose parallelism is tied to the `cycle_length` argument, the newly introduced `num_parallel_calls` argument of `tf.data.Dataset.interleave` is decoupled from the `cycle_length` argument and identifies the degree of parallelism to use for fetching output elements.
PiperOrigin-RevId: 211886816
|
| |
| |
| |
| |
| |
| | |
Estimator that is created from a Keras model.
PiperOrigin-RevId: 211886643
|
| |
| |
| |
| |
| |
| | |
Match #22072
PiperOrigin-RevId: 211884527
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Move routine ConstructIrArrayForOutputs to class IrEmitter so that it can be
used in classes IrEmitterNested and IrEmitterUnnested.
Move the code that stores the address of each individual output of a
multiple output fusion to the tuple buffer of the fusion to an overload
version of routine llvm_ir::EmitTuple so that we can reduce code duplication.
PiperOrigin-RevId: 211884483
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211883998
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211883131
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211882134
|
| |
| |
| |
| | |
PiperOrigin-RevId: 211879623
|
|\ \
| | |
| | |
| | | |
PiperOrigin-RevId: 211878263
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211876538
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211875205
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211874785
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
There is no reason for outside dependents to make a distinction between the
Eigen or MKL transpose operation, as the substitution is transparent. There is
also no need for transpose_op.cc itself to be compiled differently based on
whether MKL is in use or not. Therefore we remove external dependencies on
:mkl_transpose_op and make :transpose_op depend on it if needed (i.e., if
using MKL). This is consistent with how other transparent MKL operations (e.g.
matmul) are built.
PiperOrigin-RevId: 211874336
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211874311
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211873961
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211869673
|
| | |
| | |
| | |
| | |
| | |
| | | |
Previously, because "_Arg" nodes are considered to be "stateful", these nodes were unconditionally included in the seed set of nodes for pruning a function body. Since an "_Arg" node has no visible side effect, we can safely prune these, which makes small projection functions (like `lambda x, y: y`) more efficient.
PiperOrigin-RevId: 211867380
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 211866647
|
|\ \ \
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 211865682
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
at the runtime since in configure.py it will strip the ".0" suffix, but it makes the things cleaner and less confusing.
PiperOrigin-RevId: 211860068
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
saved_model/signature_def_utils
PiperOrigin-RevId: 211858972
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
header from generated test examples.
PiperOrigin-RevId: 211858287
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 211857599
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
individually and using consistent style for existing test names.
PiperOrigin-RevId: 211855926
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This is because many ops take longer during the first step due to autotune. Instead, the first timeline is now outputed after N seconds/steps.
PiperOrigin-RevId: 211854304
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
verifier.
PiperOrigin-RevId: 211854249
|