| Commit message (Collapse) | Author | Age |
... | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | | |
CholeskyOuterProduct.
- Deprecate event_ndims argument
- Create a Square bijector for the scalar case of CholeskyOuterProduct (which now only operates on matrices).
- Create a AffineScalar bijector for the scalar case of Affine (which now only operates on vectors)
PiperOrigin-RevId: 188801116
|
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | | |
This change harmonizes the graph construction codepaths for DirectSession and
MasterSession, which will make it easier to add new subgraph creation features.
PiperOrigin-RevId: 188799932
|
| | | | | |
|
| | | | |
| | | | |
| | | | |
| | | | | |
PiperOrigin-RevId: 188796626
|
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | | |
eager mode.
This change partially replicates the code in `tf.contrib.eager.Iterator`.
However, since that class depends on contrib-level functionality (viz.
cross-device prefetching support), we cannot move it wholesale to core.
PiperOrigin-RevId: 188790349
|
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | | |
The wrapper created by `setup.py` calls the entry point
function with no arguments. `freeze_graph.main` expects
the global `FLAGS` to be set, and one argument.
This change adds a `run_main` function to use as the entry point,
which expects no arguments and parses the flags.
It also adds a `flags` argument to `main` so the flags can be passed
directly without using a `global FLAGS` declaration.
|
| |_|/ /
|/| | |
| | | |
| | | |
| | | |
| | | |
| | | | |
- Speeds up CUDNN RNNs with fp16 input/output when possible on supported
GPUs. Computations will fall back to pseudo-fp16 if tensor op math is
not supported.
- Enabled by default, but can be disabled by setting the environment
variable TF_DISABLE_CUDNN_RNN_TENSOR_OP_MATH=1.
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188784614
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
https://www.tensorflow.org/api_docs/python/tf/multiply
PiperOrigin-RevId: 188782466
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188778072
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
to a separate stage.
1) Use a new naming scheme for optimized ops,
share it with AddOpsRewrite
2) Make sure that tests actually test that optimized
nodes exists in a graph
PiperOrigin-RevId: 188772892
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188771994
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188770522
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188766477
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188766232
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Shockingly this one was also due to PySequence_GetItem.
PiperOrigin-RevId: 188765548
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188765271
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188763442
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
- Use HloInstructionMap to get deterministic iteration order.
PiperOrigin-RevId: 188755375
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This is in preparation for enabling the C API. Modifying an op's NodeDef
directly has no effect with the C API enabled.
PiperOrigin-RevId: 188754464
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188753529
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
for `axis` arg and support taking lists.
PiperOrigin-RevId: 188751894
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
dimensions remains unchanged.
PiperOrigin-RevId: 188751074
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188750079
|
| | | |
| | | |
| | | |
| | | | |
PiperOrigin-RevId: 188747641
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
RecordMutation could race with ExtendSessionGraphHelper, which would
release the graph lock and only keep the session lock when extending
the session.
Also makes sure thread annotations are on declarations, not definitions
(otherwise they have no effect).
PiperOrigin-RevId: 188747158
|
| | | | |
|
| | | | |
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
copying the tensor.
PiperOrigin-RevId: 188742516
|
| | | | |
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
and gemmlowp. In gemmlowp the default is '1', while in eigen it is 'number of
processors'.
PiperOrigin-RevId: 188742087
|
| |/| |
|/| | | |
|
| | |/
| | |
| | |
| | | |
PiperOrigin-RevId: 188740288
|
| | |
| | |
| | |
| | |
| | |
| | | |
quantized.
PiperOrigin-RevId: 188738133
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188731961
|
| | |
| | |
| | |
| | |
| | |
| | | |
ops.
PiperOrigin-RevId: 188730560
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
consumers on a device different than themselves. They may be used to cache or route data between devices in a deliberate manner.
Simplify code in DependencyOptimizer a bit.
PiperOrigin-RevId: 188730185
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
the kernel.
With graph execution, consistency between the shapes of the arguments to
ResourceScatterUpdate is validated by the shape inference functions at graph
construction time.
With eager execution, the shape inference logic isn't executed, so inconsistent
arguments could be provided to the kernel, which would result in a segmentation
fault prior to this change. As demonstrated by the added tests.
PiperOrigin-RevId: 188729154
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188722505
|
| | |
| | |
| | |
| | |
| | |
| | | |
This potentially allows us to get rid of additional mod and div operations.
PiperOrigin-RevId: 188719238
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188698275
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188685611
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188675327
|
| | |
| | |
| | |
| | |
| | |
| | | |
Mobile net model is downloaded from tf_http_archive("tf_mobilenet") rule and renaming the asset file in assets folder has no effect.
PiperOrigin-RevId: 188672531
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188670616
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188666142
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188663018
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Previously, if the same function was being concurrently instantiated
and released:
1. Thread one could begin to instantiate the function, determine
that it already existed in the runtime, then be preempted.
2. Thread two could release the handle on the function, causing it to
be freed and removed from the `FunctionLibraryRuntime::items_` map.
3. Thread one could then incorrectly assume that the function still
existed, and fail to find it in the `FunctionLibraryRuntime::items_`
map, causing a segfault when it attempted to increment the refcount
on an uninitialized object.
PiperOrigin-RevId: 188661500
|
| | |
| | |
| | |
| | |
| | |
| | | |
compatible with XLA.
PiperOrigin-RevId: 188652533
|
| | |
| | |
| | |
| | | |
PiperOrigin-RevId: 188651655
|