Branch 178996911 #15356
Merged
Branch 178996911 #15356
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
teamdandelion
commented
Dec 14, 2017
teamdandelion
commented
- Improvement over last PR in that we rolled back a break in the CPU tests
- Fixed the same merge conflicts in tensorflow/core/platform/cloud/gcs_dns_cache.cc
- Fixed a lot of trivial merge conflicts in keras
dot(concat(..), constant) and dot(constant, concat(..)) can be rewritten to avoid the concatenate. This can itself be a win, but can also help unlock other optimization opportunities. PiperOrigin-RevId: 178691585
PiperOrigin-RevId: 178694869
PiperOrigin-RevId: 178695724
PiperOrigin-RevId: 178701096
PiperOrigin-RevId: 178703180
… errors when building externally using either the Makefile or Bazel. The macros use stderr and fprintf which may not be defined depending on the order of headers included by the .cc files. PiperOrigin-RevId: 178708839
… tensor of unknown rank as a scalar. PiperOrigin-RevId: 178710185
PiperOrigin-RevId: 178710439
There is no great need for this yet, but I noticed that the test cases were broken (they were constructing dots with unset dimension numbers), and one thing led to another. PiperOrigin-RevId: 178713597
PiperOrigin-RevId: 178715353
If the stream is not OK, the timer might not have been initialized and finalized, in which case calling timer->Nanoseconds() is illegal and will crash. PiperOrigin-RevId: 178717089
This way when a test fails, it prints out useful information about the failure, instead of "<48-byte object with these bytes: de ad be ef ...>" PiperOrigin-RevId: 178719733
… is one.
* TestUtils now supports generating random literals with more than one constraint.
There is still an error if the constraints conflict.
PiperOrigin-RevId: 178720092
PiperOrigin-RevId: 178723108
PiperOrigin-RevId: 178723711
PiperOrigin-RevId: 178724659
PiperOrigin-RevId: 178734940
…orflow#14257. PiperOrigin-RevId: 178737278
PiperOrigin-RevId: 178740804
PiperOrigin-RevId: 178751067
PiperOrigin-RevId: 178759398
1) It fixes a bug that manifested as `OutOfRange` being returned prematurely. 2) It changes the behavior on sequences of elements whose size is not a multiple of `batch_size`. Previously, the implementation would drop the last small batch (similar to `batch_and_drop_remainder). Newly, the implementation returns the last small batch (similar to `batch`). PiperOrigin-RevId: 178764508
PiperOrigin-RevId: 178767676
PiperOrigin-RevId: 178769850
Without this change, the C++ ImportGraphDef API returns unused input_map keys (which are plumbed through to the C API as well). However, the Python import_graph_def API requires slightly different semantics: it throws an error for unused input_map keys that are missing from the GraphDef. This change modifies the C and C++ APIs to limit the returned keys to those missing from the GraphDef, and plumbs this through to the C API-enabled import_graph_def implementation. Note that this is a change to the existing C API. Luckily the modified method hasn't been released yet, so it's ok to change it. PiperOrigin-RevId: 178783957
Push constants down add/mul to canonicalize chains and possibly create constant nodes at the bottom. Example:
+ + +
/ \ / \ / \
c1 + --> x + --> x c1+c2
/ \ / \
c2 x c2 c1
Small cleanup: Consolidate code for manipulating names of nodes added or modified during constant folding.
PiperOrigin-RevId: 178785218
PiperOrigin-RevId: 178787158
CompositeNodeManager has per-device LIFO manager, FirstReadyManagers for _Send and _Recv ops, and chooses FirstReady among the ops from per-device LIFOManager and _Send and _Recv FirstReadyManagers. This one can maximizes producer-consumer locality within a device (with LIFO), but does not introduce previously reported scheduling inefficiency w.r.t. multi-device execution with separately managing _Send and _Recv ops and global FirstReady policy across devices. It's implemented, but not enabled; VirtualScheduler still uses FirstReadyManager. PiperOrigin-RevId: 178787352
PiperOrigin-RevId: 178788810
PiperOrigin-RevId: 178790193
PiperOrigin-RevId: 178951330
…graph mode Fixes a bug in which EagerTensors were provided as input to an op. PiperOrigin-RevId: 178957283
- adds support for legacy "BatchMatMul" operators - adds constant scalar values to graphviz output PiperOrigin-RevId: 178957498
… new implementation will exist alongside the old one (selectable through the scheduler options) until its superiority is confirmed, at which point the old rate-based implementation will be removed. The new implementation requires fewer options and no user feedback to achieve a low latency batching. Instead of processing batches at an adjustable rate, we limit the number of batches which can be concurrently processed. Below the limit, batches are immediately processed upon creation. At the limit, the oldest batch is processed once an in-processing batch finishes. The scheduler continuously adjusts the limit in order to maintain the smallest overall latency. PiperOrigin-RevId: 178960621
PiperOrigin-RevId: 178961790
PiperOrigin-RevId: 178962340
This avoids the need for users to add `loss = loss / num_of_towers` code and is in more in line with the current best practices. I verified this by running cnn_mnist. PiperOrigin-RevId: 178963334
PiperOrigin-RevId: 178965261
Previously, Python serialization and deserialization used the half_val field of TensorProto, whereas C++ serialization used the int_val field. However, C++ bfloat16 deserialization was always broken, so it was never possible to correctly deserialize a bfloat16 Tensor. The only reason serialization worked at all was because of the generic tensor_contents bytes serialization. PiperOrigin-RevId: 178966536
…on of conditional HloInstruction. PiperOrigin-RevId: 178966782
PiperOrigin-RevId: 178966883
I had to roll in the change to generalize CPU layout assignment as without it we lose the make-rhs-column-major optimization and that causes a performance regression. PiperOrigin-RevId: 178970986
PiperOrigin-RevId: 178974641
PiperOrigin-RevId: 178977412
…ted TensorFlow. PiperOrigin-RevId: 178980799
* Remove the clustered graph part as it was difficult to keep it updated with the rest of the graph and instead operate on the graph directly; PiperOrigin-RevId: 178980836
PiperOrigin-RevId: 178984357
PiperOrigin-RevId: 178986670
PiperOrigin-RevId: 178988579
PiperOrigin-RevId: 178989673
PiperOrigin-RevId: 178995589
This allows Variants to sit on resource variables; before, though the ReadValue op was enabled for Variants on GPU, because assignment happened on CPU, Variant-based Resource Variables always had to reside on CPU due to the associated colocation constraints. PiperOrigin-RevId: 178996911
gunan
approved these changes
Dec 14, 2017
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.