| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 199862180
|
|
|
|
| |
PiperOrigin-RevId: 199859711
|
|
|
|
|
|
| |
function.
PiperOrigin-RevId: 199859228
|
|
|
|
|
|
| |
Closes #19808.
PiperOrigin-RevId: 199857219
|
|
|
|
| |
PiperOrigin-RevId: 199855838
|
|
|
|
|
|
| |
arrays for dilated unstrided 1x1 cases.
PiperOrigin-RevId: 199849200
|
|
|
|
| |
PiperOrigin-RevId: 199846447
|
|
|
|
| |
PiperOrigin-RevId: 199839745
|
|
|
|
| |
PiperOrigin-RevId: 199836336
|
|
|
|
|
|
|
|
|
|
| |
The new HLO instruction serves two purposes. (1) It generates a new token value. This is the only way to create tokens. (2) The operation is variadic, taking zero or more token operands. The operation acts as a join of its operands.
I considered initially using a kConstant constant as a method to create new tokens, but this ran into problems because of expectations in backends regarding constants and their materialization.
This CL enables creation of generate-token instructions, but the new instruction is not supported yet in any backend.
PiperOrigin-RevId: 199836205
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fix for: #17315
If numpy array is read-only, calling PyArray_FromAny with NPY_ARRAY_CARRAY
flags introduce extra memory copy.
Before:
feed_cpu_variable_read_only: 5.6 GB/sec, min: 17.99, median: 19.54, mean: 19.76
After:
feed_cpu_variable_read_only: 13.2 GB/sec, min: 7.60, median: 7.78, mean: 8.13
PiperOrigin-RevId: 199835695
|
|
|
|
|
|
| |
`shuffle_and_repeat(...)`.
PiperOrigin-RevId: 199834400
|
|
|
|
| |
PiperOrigin-RevId: 199834091
|
|
|
|
| |
PiperOrigin-RevId: 199833632
|
|
|
|
|
|
|
|
| |
* Mention both keras and tf.keras
* In one of the early paragraphs, list all three high-level APIs supported
(tf.estimator, keras and tf.contrib.slim).
PiperOrigin-RevId: 199830255
|
|
|
|
| |
PiperOrigin-RevId: 199826723
|
|
|
|
|
|
| |
pop, stack. Simplify the type annotation mechanism by having it literally copy its arguments, instead of attempting to resolve them.
PiperOrigin-RevId: 199822771
|
|
|
|
| |
PiperOrigin-RevId: 199822232
|
|
|
|
| |
PiperOrigin-RevId: 199821675
|
|
|
|
|
|
| |
models.
PiperOrigin-RevId: 199817660
|
|
|
|
|
|
|
| |
convolution_test had a zero relative error bound which made it overly sensitive
to changes to the underlying computation.
PiperOrigin-RevId: 199814523
|
|
|
|
| |
PiperOrigin-RevId: 199809082
|
|
|
|
|
|
| |
Stop creating pre-fused nodes in BatchNormExpander.
PiperOrigin-RevId: 199807585
|
|
|
|
| |
PiperOrigin-RevId: 199807061
|
|
|
|
|
|
|
|
|
| |
Previously, we accepted the `padded_shapes` argument without validating that
it was compatible with the `input_dataset.output_shapes`. In many cases, we have
enough static shape information to do this, and so we now raise an actionable
error at the point where the mistake is committed, rather than at runtime.
PiperOrigin-RevId: 199800348
|
|
|
|
| |
PiperOrigin-RevId: 199795176
|
|
|
|
| |
PiperOrigin-RevId: 199782188
|
|
|
|
| |
PiperOrigin-RevId: 199780350
|
|
|
|
|
|
| |
inputs
PiperOrigin-RevId: 199776409
|
|
|
|
|
|
|
| |
There's no real reason to not support queues in eager for people using them
without using queue runners.
PiperOrigin-RevId: 199770626
|
|
|
|
|
|
| |
With this change, it is now possible to embed differentiable eager code --- running on either CPU or GPU --- in graphs. Higher-order derivatives are not yet supported.
PiperOrigin-RevId: 199768301
|
|
|
|
| |
PiperOrigin-RevId: 199765487
|
|
|
|
|
|
| |
This allows checkpointing input pipelines with .shuffle(reshuffle_each_iteration=True[default]) and .list_files().
PiperOrigin-RevId: 199753836
|
|
|
|
|
|
|
|
| |
fixes in the patch have been incorporated into the Eigen opensource repository with this commit:
https://bitbucket.org/eigen/eigen/commits/60ab50654998f1cbe2791d49fea94d0ca5ae08a8
PiperOrigin-RevId: 199749536
|
|
|
|
| |
PiperOrigin-RevId: 199745064
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The current implementation tiles the x-dimension of the tensors to calculate the
partial results of the reduction. This change increases such an x-tile size from
8 to 64 if doing so results in all saturated tiles. Otherwise, this change adds
z-dimension tiles to increase the number of elements that each thread reduces to
a partial result to reduce the number of needed dynamic atomic operations and
intra-warp reduction operations.
Use a tighter yet safe loop bound for the last unsaturated tile.
Avoid generating the atomic operation when the tile size is not smaller than the
reduction width.
Extend ForLoop emitter to support a request for fully loop unrolling.
Add three tests.
PiperOrigin-RevId: 199744209
|
|
|
|
|
|
| |
Make it easier for analysis and code search.
PiperOrigin-RevId: 199736646
|
|
|
|
| |
PiperOrigin-RevId: 199735191
|
|
|
|
|
|
|
|
|
|
| |
buffer.
- Previously, we say we cannot reuse operand buffer for a loop fusion node if any of the fusion's inputs is a broadcast or reshape. That's too conservative since in theory we can still reuse the operand's buffer if all the users of that particular operand are elementwise. This CL implements that.
- Also fixed a bug in previous code where a dynamic update fusion node that ends with convert (added for bf16) is not caught by the if condition currectly.
PiperOrigin-RevId: 199731488
|
|
|
|
|
|
|
|
| |
BatchFunction op.
o Renames the previous version to batch_function_v1.
PiperOrigin-RevId: 199729701
|
|
|
|
| |
PiperOrigin-RevId: 199729533
|
|
|
|
|
|
| |
even in the case of falsy eval_results
PiperOrigin-RevId: 199728382
|
|
|
|
| |
PiperOrigin-RevId: 199726426
|
|
|
|
|
|
| |
Slice doesn't read the entire input. It only reads enough to make the output.
PiperOrigin-RevId: 199722987
|
|
|
|
| |
PiperOrigin-RevId: 199722844
|
|
|
|
|
|
|
|
|
|
|
| |
See #19657 for some motivation.
Without this explicit check, a large graph would trigger an assertion failure
in the protobuf codebase
(https://github.com/google/protobuf/blob/0456e269ee6505766474aa8d7b8bba7ac047f457/src/google/protobuf/message_lite.cc#L68)
Pull Request for google/protobuf: https://github.com/google/protobuf/pull/4739
PiperOrigin-RevId: 199719082
|
|
|
|
|
|
|
|
| |
statistics from `Example` record of dataset.
This change-list also applies transformation function `feature_stats()` to collect stats in an associated stats_aggregator (if any) to dataset in `make_batched_feature_dataset()` by default.
PiperOrigin-RevId: 199718439
|
|
|
|
| |
PiperOrigin-RevId: 199714511
|
|
|
|
|
|
| |
HloInstruction.
PiperOrigin-RevId: 199712253
|
|
|
|
| |
PiperOrigin-RevId: 199706732
|