- 23 2月, 2019 40 次提交
-
-
由 Dan Moldovan 提交于
Fix rewrap to always return the updated entity. Use it as such at the autograph integration point. In turn, temporarily disable autograph for dataset functions until we reduce the conversion overhead. Fixes #25159. With thanks to @ktaebum for identifying this fix. PiperOrigin-RevId: 235255035
-
由 Scott Zhu 提交于
PiperOrigin-RevId: 235252398
-
由 Nick Felt 提交于
This changes the resource model for the TF 2.0 tf.summary.SummaryWriter as returned by tf.summary.create_file_writer(). Previously, we preserved the 1.x tf.contrib.summary.create_file_writer() semantics where underlying resources were shared by logdir to facilitate reuse of a single eventfile per directory, which avoids accidentally overwriting an existing eventfile (which is possible if both are created within the same second) and is better supported by TensorBoard. Unfortunately, that sharing behavior resulted in buggy and confusing resource deletion behavior as described in #25707, and is also out of alignment with the overall TF 2.0 philosophy that resource lifetimes should match the lifetimes of their encapsulating Python objects. The new SummaryWriter wrapper object fixes this by generating a unique resource name for each instance. Inside @tf.functions, this requires lifting initialization out using init_scope(), which means in-graph tensors can no longer be passed to create_file_writer(), and the returned SummaryWriter must be stored in a non-local variable so that it outlives the tracing phase and is still alive when the function graph executes. A future change will add better error messaging for the latter circumstance. To keep the SummaryWriter <-> resource mapping 1:1, this also changes the close() method so that in eager mode it will permanently close the SummaryWriter and prevent re-initialization. (A future change may do this for graph mode as well.) To mitigate the risk of eventfile name collisions, the new wrapper bakes the PID and a TF-generated UID into the filename suffix. Finally, this cleans up the public API surface by 1) moving the "logdir=None means no-op" behavior into a dedicated create_noop_writer() factory function and 2) hiding SummaryWriter.__init__() from the public API by making SummaryWriter an abstract base class; instances should be obtained via factory functions anyway. Fixes #25707. A subsequent change will update the TF 2.0 Keras callback to call create_file_writer_v2(), which should address #25524 and #24632. PiperOrigin-RevId: 235249883
-
由 Allen Lavoie 提交于
PiperOrigin-RevId: 235247680
-
由 Allen Lavoie 提交于
This is extra "nice to have" metadata, and otherwise it looks like a checkpointing error. Not worth bothering people about. PiperOrigin-RevId: 235247436
-
由 Justin Lebar 提交于
These are subtly different: - sqrt(-inf) = nan, but pow(-inf, 0.5) = +inf Add exhaustive tests for sqrt, pow(0.5), and move the exhaustive fp16 test for log into the new file. Doing this requires a nontrivial rewrite of the exhaustive test; hopefully this is the last time. While we're here, we enable a few more exhaustive tests. Also remove the special-case handling in GPU elemental IR emitter for pow(x, 0.5) now that we have a sqrt HLO. PiperOrigin-RevId: 235245101
-
由 Rohan Jain 提交于
Splitting out the memory usage tests into their own test file as that was making the test suite for MultiDeviceIterator too big and timing out. PiperOrigin-RevId: 235242183
-
由 Pavithra Vijay 提交于
PiperOrigin-RevId: 235241120
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235238247
-
由 Rick Chao 提交于
Remove early_stopping.py from estimator/contrib in favor of estimator/python/estimator/early_stopping.py. And the test. PiperOrigin-RevId: 235236119
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 235235216
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 235233898
-
由 Sanjoy Das 提交于
PiperOrigin-RevId: 235233320
-
由 William Chargin 提交于
In cl/234179031, we improved this test to use `os.path.join` rather than a hard-coded `+ "/tb"` in a few places, but apparently missed some. Tested: After 1036s/tb/something_else/, tests still pass on my local gLinux desktop. Before this change, tests would fail after that substitution, albeit not for the same reason as the traceback in b/125762969: we get a normal `assertItemsEqual` failure rather than an ENOENT. Running Kokoro Windows presubmits per go/tf-testing passes. RELNOTES: n/a PiperOrigin-RevId: 235232824
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235232101
-
由 Pavithra Vijay 提交于
PiperOrigin-RevId: 235229813
-
由 Frank Chen 提交于
PiperOrigin-RevId: 235229164
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235228999
-
由 Derek Murray 提交于
PiperOrigin-RevId: 235228690
-
由 Francois Chollet 提交于
PiperOrigin-RevId: 235226794
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235226607
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235226116
-
由 Pavithra Vijay 提交于
- Disable `run_eagerly` and distribution strategy if there are symbolic tensors added to the model using `add_metric` or `add_loss`. - Added error message for when subclassed model/Sequential model without input_shape is used with DS. - Add error message when `run_eagerly` is enabled with DS. PiperOrigin-RevId: 235225004
-
由 James Ring 提交于
PiperOrigin-RevId: 235221008
-
由 Gunhan Gulsoy 提交于
PiperOrigin-RevId: 235219008
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 235218660
-
由 Derek Murray 提交于
Previously, the tuple would be unpacked in string formatting, and if it had more than one element we would get an argument unpacking error, instead of the intended TypeError. PiperOrigin-RevId: 235217943
-
由 Derek Murray 提交于
Internal users of these properties should now use dataset_ops.get_legacy_output_{shapes,types,classes}() instead. In a future CL we will remove the properties from the DatasetV2. PiperOrigin-RevId: 235217553
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235215099
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 235214785
-
由 A. Unique TensorFlower 提交于
Provides better memory locality which speeds up most ops, and makes destruction trivial. PiperOrigin-RevId: 235213734
-
由 Allen Lavoie 提交于
An unfortunate combination of the name_scope method wrapping not forwarding argspecs and input_signatures inspecting functions for keyword arguments. PiperOrigin-RevId: 235211265
-
由 A. Unique TensorFlower 提交于
Notice that most of the constants created by the constant folding pass are already in compressed format. This just extends this practice to all constants in the graph that have at least 64 elements and where the achievable compression ratio is grater or equal to 2. PiperOrigin-RevId: 235211073
-
由 Bruce Fontaine 提交于
PiperOrigin-RevId: 235210894
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235210508
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 235207900
-
由 Zhenyu Tan 提交于
PiperOrigin-RevId: 235207665
-
由 Rohan Jain 提交于
Renames to prepare for the TF2.0 API. HashTable -> StaticHashTable and MutableDenseHashTable -> DenseHashTable. Also insert -> insert_or_assign and remove -> erase for DenseHashTable. PiperOrigin-RevId: 235206233
-
由 Guangda Lai 提交于
PiperOrigin-RevId: 235201217
-
由 Dan Moldovan 提交于
PiperOrigin-RevId: 235197577
-