- 12 7月, 2019 20 次提交
-
-
由 William Chargin 提交于
Now that TensorBoard 1.14.0 has been released, our nightlies have moved up to 1.15.x alphas. PiperOrigin-RevId: 257662026
-
由 Peter Hawkins 提交于
PiperOrigin-RevId: 257660458
-
由 Penporn Koanantakool 提交于
PiperOrigin-RevId: 257656880
-
由 Guangda Lai 提交于
Make TF-TRT python tests 2.0 compatible by adding v2 only tests and decorating all existing tests with run_v1_only. PiperOrigin-RevId: 257653149
-
由 Guangda Lai 提交于
since GraphConverter is not used elsewhere. PiperOrigin-RevId: 257653003
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 257650763
-
由 Pavithra Vijay 提交于
Fix issue - in multi-output model if class weight is specified just for some outputs the weights get incorrectly mapped to placeholders PiperOrigin-RevId: 257650645
-
由 A. Unique TensorFlower 提交于
This allows for the attribute to hold symbolic references to other operations than FuncOp. This also allows for removing the dependence on FuncOp from the base Builder. PiperOrigin-RevId: 257650017
-
由 Andy Ly 提交于
PiperOrigin-RevId: 257648888
-
由 A. Unique TensorFlower 提交于
Occasionally tensorflow::HistogramFixedWidthFunctor::Compute triggers a segmentation fault 11. It happens when the argument i of index_to_bin() is negative, which is caused by casting a big (x-a)/step int64 value that exceeds int32 range to int32. Switching the order of cwiseMin and cast fixes the problem. PiperOrigin-RevId: 257647164
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 257646455
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 257644752
-
由 Ashwin Murthy 提交于
PiperOrigin-RevId: 257644235
-
由 Sanjoy Das 提交于
//tools/target_cpu:haswell isn't present in open source. PiperOrigin-RevId: 257644093
-
由 Derek Murray 提交于
This change replaces an std::function<> instance variable that has the same value for every KernelAndDevice object with a static local variable. PiperOrigin-RevId: 257641403
-
由 A. Unique TensorFlower 提交于
Add detection_responder which allows each platform to process the person detection output in its own way. For example, sparkfun_edge lights up the yellow LED for no person and the green LED for person, and toggles the blue LED on each run. PiperOrigin-RevId: 257638242
-
由 Gaurav Jain 提交于
PiperOrigin-RevId: 257637727
-
由 Ihor Indyk 提交于
[tf.data] Increase the number of `iterator.get_next()` calls to 10000 (from 1000) in `map_and_interleave` autotuning benchmark to decrease variance of the output. PiperOrigin-RevId: 257624991
-
由 Prakalp Srivastava 提交于
If the return of a function is an unsigned tensor type, it was being modeled as standard int type, losing information that the return is unsigned. PiperOrigin-RevId: 257623093
-
由 TensorFlower Gardener 提交于
Merge pull request #30299 from ROCmSoftwarePlatform:google_upstream_rocm_stream_executor_updates_190701 PiperOrigin-RevId: 257617613
-
- 11 7月, 2019 20 次提交
-
-
由 Sergei Lebedev 提交于
PiperOrigin-RevId: 257614950
-
由 Peter Hawkins 提交于
Also move RegisterCpuCustomCallTarget into the main Python binding module (xla.cc) since it is unrelated to anything else in local_client.*. PiperOrigin-RevId: 257612668
-
由 Edward Loper 提交于
In NestedStructureCoder: fixed bug where a TensorSpec with name=None would get deserialized with name=''. PiperOrigin-RevId: 257610414
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 257610259
-
由 Peter Hawkins 提交于
Currently on CPU and GPU there is no limit to how many operations the host can enqueue on the device stream. On GPU this doesn't usually cause a problem because the allocator is logically synchronized to the tail of the compute stream, and so we can free and reuse memory for operations enqueued on the stream. On CPU, the allocator is logically synchronized to the head of the compute stream, which means that the allocator cannot reuse buffers between operations enqueued on the stream. This means that the memory usage is proportional to the number of enqueued operations, which can rapidly blow up. Add a semaphore class and use it to set a moderate limit on the depth of the queue (32). The existing "synchronous" mode, used on TPU at present, is a special case of this support where the queue depth is 1. This may help with https://github.com/google/jax/issues/928 . PiperOrigin-RevId: 257606960
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 257605915
-
由 Doe Hyun Yoon 提交于
The way VirtualScheduler currently checks node ready is to count how many input nodes are ready; it's possible that a certain pattern (e.g., Switch - Merge) can break this assumption; both outputs of Switch is triggered, and then Merge becomes ready twice. If a Merge node is executed more than once, we skip triggering its output node. PiperOrigin-RevId: 257603202
-
由 Edward Loper 提交于
PiperOrigin-RevId: 257597675
-
由 Peter Hawkins 提交于
Consolidate on the variant that calls destructors from a worker thread; of the two uses of the non-threaded version, one was buggy since it might have freed GPU buffers, and for the other the benefit seemed minimal. Also remove overly complicated and confusing code that tried to control which thread released a reference; the only thing that matters is that on GPU references to GPU memory aren't released from a device callback. Change PythonRefManager so the RAII class it returns is non-copyable but is wrapped in a std::shared_ptr<>. This makes its copying semantics easier to reason about. PiperOrigin-RevId: 257597653
-
由 Sergei Lebedev 提交于
PiperOrigin-RevId: 257596912
-
由 TensorFlower Gardener 提交于
PiperOrigin-RevId: 257595934
-
由 Edward Loper 提交于
PiperOrigin-RevId: 257595324
-
由 Sergei Lebedev 提交于
There is no reason to create a new status on each call. PiperOrigin-RevId: 257594441
-
由 Peter Hawkins 提交于
[XLA:Python] Remove some special cases in the XLA Python bindings that exist to work around the fact that the Host device didn't support events. Enable "multi stream" mode everywhere and delete the code for "single stream" mode. This has the side effect of using more threading than we did previously on CPU, in particular the "host to device" and "device to host" transfers are done in separate streams (threads). It's not clear that's a bad thing, and more commonality between backends is nice. [SE:Host] Implement the se::Event API for the host platform, as a minimal wrapper around absl::Notification. PiperOrigin-RevId: 257594142
-
由 Peter Hawkins 提交于
On GPU, events are expensive to create and apparently can and should be reused. PiperOrigin-RevId: 257591195
-
由 Adrian Kuegel 提交于
Everywhere else we already check that the fusions we create will not become too big. We need to check it in fusion_merger as well. PiperOrigin-RevId: 257589636
-
由 Ian Langmore 提交于
`convert_nonref_to_tensor` PiperOrigin-RevId: 257584511
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 257568311
-
由 A. Unique TensorFlower 提交于
PiperOrigin-RevId: 257568308
-
由 Sergei Lebedev 提交于
The precise contract for conversion functions is currently undocumented, but the implementation additionally allowed errors.UnimplementedError and errors.InvalidArgumentError to be raised. As far as I can tell, these are not required by any existing internal conversion function. PiperOrigin-RevId: 257567409
-