| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 159902919
|
|
|
|
| |
PiperOrigin-RevId: 159901417
|
|
|
|
| |
PiperOrigin-RevId: 159899925
|
|
|
|
|
|
|
|
|
|
|
| |
* DNNClassifier
* DNNRegressor
* LinearClassifer
* LinearRegressor
* DNNLinearCombinedClassifier
* DNNLinearCombinedRegressor
PiperOrigin-RevId: 159898954
|
|
|
|
|
|
| |
Add expensive operation checker
PiperOrigin-RevId: 159897279
|
|
|
|
| |
PiperOrigin-RevId: 159896013
|
|
|
|
|
|
|
| |
https://reviews.llvm.org/rL305938). Disable TBAA (to stop the miscompiles)
while we fix the root issue.
PiperOrigin-RevId: 159895736
|
|
|
|
| |
PiperOrigin-RevId: 159886125
|
|
|
|
|
|
|
|
| |
Because some higher level APIs continue to use the legacy name format,
when using ClusterSpec propagation, we need to ensure that we register
the devices under their legacy names as well as their canonical names.
PiperOrigin-RevId: 159885777
|
|
|
|
| |
PiperOrigin-RevId: 159883986
|
|
|
|
|
|
|
|
| |
everyone.
Also fix a bug in the C++ op when the input ST has 0 elements.
PiperOrigin-RevId: 159880044
|
|
|
|
| |
PiperOrigin-RevId: 159879840
|
|
|
|
|
|
| |
ValueOrDie.
PiperOrigin-RevId: 159879754
|
|
|
|
|
|
|
| |
the reparameterization and score-gradient tricks and does not depend on
tf.Distribution like inputs.
PiperOrigin-RevId: 159877923
|
|
|
|
| |
PiperOrigin-RevId: 159876942
|
|
|
|
|
|
| |
This CL makes HloModule clonable, which is necessary when we want to run the same compilation twice with the same input.
PiperOrigin-RevId: 159874256
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Ensure that the max number of categories does not exceed largest integer-form float.
- Make dtype inference consistent between Categorical and Multinomial
distributions.
- Improve documentation to better reflect that the Categorical
distribution is analogous to `argmax{OneHotCategorical}` (itself being
identical to `argmax{Multinomial(p,n=1)}` but not Multinomial.
- Fix validation_args Heisenberg uncertainty: only validation logic should live under self.validate_args. E.g., validate_args=True would sometimes imply `x=floor(x)` which changes behavior thus making debugging impossible because enabling validation *changes* values.
- Corrected `Geometric` swapping of validate_args` and `allow_nan_stats` default-values.
Fixes #10149
PiperOrigin-RevId: 159872532
|
|
|
|
|
|
|
| |
building the Graph. This allows loading a graph that contains ops not
needed for the compiled binary.
PiperOrigin-RevId: 159869692
|
|
|
|
|
|
| |
I didn't use that.
PiperOrigin-RevId: 159869169
|
|
|
|
| |
PiperOrigin-RevId: 159855254
|
|
|
|
|
|
| |
the sharded embeddings.
PiperOrigin-RevId: 159854706
|
|
|
|
|
|
| |
optimized away
PiperOrigin-RevId: 159853171
|
|
|
|
| |
PiperOrigin-RevId: 159852889
|
|
|
|
|
|
| |
avoid incorrect threading test failures (tsan).
PiperOrigin-RevId: 159852818
|
|
|
|
| |
PiperOrigin-RevId: 159852557
|
|
|
|
|
|
| |
static shape is not present.
PiperOrigin-RevId: 159852297
|
|
|
|
|
|
|
| |
The batch size is configured as follows:
The user may specify a global batch size in their hyperparameters. If the 'batch_size' field is set, then we convert the global batch size into a per-shard batch size by dividing by num_shards before running their input_fn.
PiperOrigin-RevId: 159851773
|
|
|
|
| |
PiperOrigin-RevId: 159850619
|
|
|
|
| |
PiperOrigin-RevId: 159849759
|
|
|
|
| |
PiperOrigin-RevId: 159847551
|
|
|
|
|
|
| |
Made name_space usage consistent.
PiperOrigin-RevId: 159846928
|
|
|
|
| |
PiperOrigin-RevId: 159841935
|
|
|
|
| |
PiperOrigin-RevId: 159840822
|
|
|
|
| |
PiperOrigin-RevId: 159835723
|
|
|
|
| |
PiperOrigin-RevId: 159820487
|
|
|
|
|
|
|
|
| |
a pool to be a global pool. This allows a division
between large and small pools, without needing to make
new pool for each session.
PiperOrigin-RevId: 159789678
|
|
|
|
| |
PiperOrigin-RevId: 159789178
|
|
|
|
| |
PiperOrigin-RevId: 159787178
|
|
|
|
| |
PiperOrigin-RevId: 159781607
|
|
|
|
|
|
|
|
|
| |
Assigning __defaults__ relies on an implementation detail of CPython, confuses
type checkers (and developers :)), and is error-prone since it doesn't make the
relationship between parameter names and default values explicit.
This CL switches to overloading __new__ instead.
PiperOrigin-RevId: 159773922
|
|
|
|
| |
PiperOrigin-RevId: 159772841
|
|
|
|
|
|
| |
functions, memory corruptions, ...)
PiperOrigin-RevId: 159772801
|
|
|
|
| |
PiperOrigin-RevId: 159771215
|
|
|
|
| |
PiperOrigin-RevId: 159769520
|
|
|
|
| |
PiperOrigin-RevId: 159764956
|
|
|
|
| |
PiperOrigin-RevId: 159763112
|
|
|
|
|
|
|
|
| |
Change is made in a backwards and forward compatible manner, since
we add a new attribute with a default that remains the same, and
simply register a few new kernels.
PiperOrigin-RevId: 159761347
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Add monotonic_attention and safe_cumprod helper functions.
* Add _BaseMonotonicAttentionMechanism base class.
* Add BahdanauMonotonicAttention and LuongMonotonicAttention classes.
These attention mechanisms are proposed in
Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, Douglas Eck,
"Online and Linear-Time Attention by Enforcing Monotonic Alignments."
ICML 2017. https://arxiv.org/abs/1704.00784
PiperOrigin-RevId: 159760073
|
|
|
|
|
|
|
|
| |
Remove Service::runs_in_client_process_ field and it's dead user. This was
previously used by the "InProcess" methods which have been replaced with
the LocalClient API.
PiperOrigin-RevId: 159759455
|
|
|
|
| |
PiperOrigin-RevId: 159757904
|