| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 182127737
|
|
|
|
|
|
|
|
| |
resolution. Previously, the input resolution was not taken into account, which led to undefined padding for many well-known models (since those rely on SAME padding, and in some cases SAME padding depends on input resolution).
This change also redesigns many aspects of the topological sorting and layer parsing functions, introducing new modules and tests.
PiperOrigin-RevId: 182124694
|
|
|
|
| |
PiperOrigin-RevId: 182124532
|
|
|
|
|
|
| |
Users of weighted losses need loss_reduction=SUM, otherwise the most common choice is SUM_BY_NONZERO_WEIGHTS. Estimator class doesn't have visibility into whether weighted losses are used, so the default was assumed to be SUM, because it's correct-ish in either case. This change sets the default to SUM_BY_NONZERO_WEIGHTS to accomodate the most common case.
PiperOrigin-RevId: 182122737
|
|
|
|
|
|
|
|
| |
I'm not totally sure why this was done to begin with. However, nothing
seems to break with it removed. I'd like to remove this since setting
_shape won't work with the C API enabled.
PiperOrigin-RevId: 182122350
|
|
|
|
| |
PiperOrigin-RevId: 182121746
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
throws an exception in tf.name_scope() and tf.variable_scope().
The previous conversion from generator style context managers translated a
nested with block like
with context_a, context_b:
yield
into the following pair of __enter__()/__exit__() methods:
def __enter__(self):
self.context_a.__enter__()
self.context_b.__enter__()
def __exit__(self, *args):
self.context_b.__exit__(*args)
self.context_a.__exit__(*args)
return False
This translation is only correct when self.context_b.__enter__() does not throw
an exception. In the context of tf.name_scope() and tf.variable_scope(),
context_a is the default graph stack, and context_b is the actual scope to
enter. Entering an actual scope throws a ValueError when the scope name is
invalid. In that case, the above implementation leaves __enter__() without
popping the default graph stack. Sub-sequent calls to pop the stack will thus
fail, obscuring the actual exception that was raised in entering the scope.
PiperOrigin-RevId: 182119816
|
|
|
|
| |
PiperOrigin-RevId: 182119760
|
|
|
|
|
|
|
| |
Fixes a bug that arose when the supplied target to `make_decorator` was
a callable that did not have a `__name__` member.
PiperOrigin-RevId: 182119694
|
|
|
|
|
|
| |
Also actively reset input_impl_ when it is exhausted.
PiperOrigin-RevId: 182118637
|
|
|
|
| |
PiperOrigin-RevId: 182112914
|
|
|
|
| |
PiperOrigin-RevId: 182112167
|
|
|
|
| |
PiperOrigin-RevId: 182111798
|
|
|
|
|
|
|
|
| |
This new argument allows function instantiators to provide a unique
string that identifies the state that is associated with the
instantiated function (i.e. typically in stateful kernels).
PiperOrigin-RevId: 182110020
|
|
|
|
|
|
| |
running it with GPU config.
PiperOrigin-RevId: 182102413
|
|
|
|
| |
PiperOrigin-RevId: 182100616
|
|
|
|
|
|
| |
Fixes #16160
PiperOrigin-RevId: 182098964
|
|
|
|
| |
PiperOrigin-RevId: 182098537
|
|
|
|
| |
PiperOrigin-RevId: 182098104
|
|
|
|
| |
PiperOrigin-RevId: 182094260
|
|
|
|
|
|
| |
only turned on by default for models without feed nodes.
PiperOrigin-RevId: 182090076
|
|
|
|
|
|
|
| |
This isn't necessary for anything, but it limits the surface area of
private API abuse.
PiperOrigin-RevId: 182086883
|
|
|
|
|
|
| |
This prevents issues like https://github.com/bazelbuild/bazel/issues/4425
PiperOrigin-RevId: 182085505
|
|
|
|
| |
PiperOrigin-RevId: 182084336
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(1) Adds a function `make_template_internal` that takes an optional keyword
argument `create_graph_function_` that controls whether or not `func_` is
compiled into and executed as a graph function.
(2) Exposes `make_template_internal` as `tfe.make_template`, so users may
write something like
`
compiled = tfe.make_template("my_op", func, create_graph_function_=True)
`
to obtain a templatized version of func that is executed as a graph
function.
(3) Simplifies the implementation of (Eager)Template's _call_func and __call__
methods in a minor way.
PiperOrigin-RevId: 182082866
|
|
|
|
| |
PiperOrigin-RevId: 182080904
|
|
|
|
|
|
| |
a control structure. The type of such object cannot be reliably inferred.
PiperOrigin-RevId: 182072568
|
|
|
|
| |
PiperOrigin-RevId: 182072485
|
|
|
|
| |
PiperOrigin-RevId: 182071549
|
|
|
|
|
|
| |
input/output array optimizations.
PiperOrigin-RevId: 182063294
|
|
|
|
|
|
| |
tuple shardings were not.
PiperOrigin-RevId: 182043879
|
|
|
|
| |
PiperOrigin-RevId: 182039394
|
|
|
|
|
|
|
| |
_set_device now works with the C API enabled, so we no longer need
this explicit test coverage.
PiperOrigin-RevId: 182014804
|
|
|
|
| |
PiperOrigin-RevId: 182007077
|
|
|
|
| |
PiperOrigin-RevId: 182005780
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
destroy the local client instance.
Add a user defined move constructor for KernelBase to nullify the parent stream
pointer. This is needed to avoid unloading kernels that haven't been loaded
when the moved-from KernelBase objects are destructed.
Add a call to ClientLibrary::DestroyLocalInstances to destroy the local client
instance used by replay_computation.
PiperOrigin-RevId: 181995818
|
|
|
|
| |
PiperOrigin-RevId: 181993953
|
|
|
|
| |
PiperOrigin-RevId: 181976155
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, DeleteWorkerSession was responsible for freeing the
WorkerSession owned by the SessionMgr. However, it is possible for
other requests to be in-flight on the same session, and requests from
the master to be delivered out of order, which leads to the potential
for a request to use a WorkerSession after it has been freed. Revise
the SessionMgr interface to handle std::shared_ptr<WorkerSession>
instead of raw pointers to avoid this risk.
PiperOrigin-RevId: 181975078
|
|
|
|
| |
PiperOrigin-RevId: 181943238
|
|
|
|
|
|
| |
tf.contrib.lookup.LookupInterface.lookup.
PiperOrigin-RevId: 181928781
|
|
|
|
|
|
|
|
| |
loaded SaveModelBundle.
#14567
PiperOrigin-RevId: 181887870
|
|
|
|
|
|
|
|
| |
mechanism implemented in
tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py
PiperOrigin-RevId: 181867510
|
|
|
|
| |
PiperOrigin-RevId: 181865795
|
|
|
|
|
|
| |
between -1 and 1.
PiperOrigin-RevId: 181864303
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Dump a log of the allocator's internal state on allocation failure at
vlog level 2 and above.
- When printing the allocator's chunks, print free chunks inline with
occupied chunks. This way fragmentation is easier to see.
- Add some GUARDED_BY annotations.
- Use a proper atomic for the "non-fatal OOM" log counter.
PiperOrigin-RevId: 181851550
|
|
|
|
|
|
| |
analyzer to better match Python semantics.
PiperOrigin-RevId: 181847178
|
|
|
|
| |
PiperOrigin-RevId: 181831552
|
|
|
|
| |
PiperOrigin-RevId: 181831249
|
|
|
|
| |
PiperOrigin-RevId: 181827605
|