| Commit message (Collapse) | Author | Age |
|
|
|
| |
Change: 149048750
|
|
|
|
| |
Change: 149047908
|
|
|
|
|
|
| |
when batch size is 1. This resulted in reducing the wall time spent in the op by 42% during testing on a Pixel phone.
Change: 149047766
|
|
|
|
|
| |
C++ microbenchmarks results are now normalized by iters (like python benchmarks).
Change: 149045367
|
|
|
|
| |
Change: 149043862
|
|
|
|
| |
Change: 149040924
|
|
|
|
| |
Change: 149040750
|
|
|
|
|
|
| |
element to contain wildcards.
Before we did wildcard expansion only if a single element is given.
Change: 149039100
|
|
|
|
|
|
| |
errors.
Change: 149038920
|
|
|
|
| |
Change: 149036604
|
|
|
|
|
| |
See https://www.tensorflow.org/api_docs/python/tf/nn/conv2d, where the "Returns" section is incomplete: Conv2D does return a Tensor that "Has the same type as Input", but it doesn't discuss what the the explicit dimensionality is (4-D), or what the output shape ([h, w, d1, d2]) is.
Change: 149036536
|
|
|
|
| |
Change: 149028834
|
|
|
|
| |
Change: 149025947
|
|
|
|
|
|
| |
1) fix bug where command triggered by mouse gets truncated if length exceeds command input box width
2) add waiting toast message so that the user knows TFDBG is crunching number on time-consuming operations.
Change: 149022490
|
|
|
|
| |
Change: 149021831
|
|
|
|
|
|
| |
imaginary part. Fix a bug in test util function assertAllCloseAccordingToType, which wasn't picking up the right values for complex64.
Change: 149016078
|
|
|
|
|
|
| |
constructor.
Change: 149004675
|
|
|
|
|
|
| |
Using name_scope means self.name = the_initialization_name + "/", which makes
it hard to set a distribution name, then know what the name is later.
Change: 148995515
|
|
|
|
| |
Change: 148970303
|
|
|
|
| |
Change: 148967694
|
|
|
|
| |
Change: 148962450
|
|
|
|
| |
Change: 148960531
|
|
|
|
|
| |
Use os.path.normpath to normalize backslashes in paths to forward slash before comparing paths.
Change: 148960384
|
|
|
|
| |
Change: 148959101
|
|
|
|
| |
Change: 148954762
|
|
|
|
| |
Change: 148954491
|
|
|
|
| |
Change: 148952814
|
|
|
|
|
| |
While we are here, add support for getting the cost analysis for call HLOs.
Change: 148952748
|
|
|
|
|
| |
computing mean().
Change: 148951756
|
|
|
|
|
| |
Updated the demo to use only local data files relevant to the demo. Also made the demo an iron snippet so folks can see the code next to the demo.
Change: 148950452
|
|
|
|
| |
Change: 148947675
|
|
|
|
| |
Change: 148945600
|
|
|
|
| |
Change: 148942856
|
|
|
|
|
|
|
| |
call the parent scope's custom_getter instead of overriding it. The child-most
scope passes the "true" VariableScope getter all the way through to the
parent-most getter for the very innermost variable access.
Change: 148940536
|
|
|
|
| |
Change: 148939552
|
|
|
|
| |
Change: 148936943
|
|
|
|
| |
Change: 148934142
|
|
|
|
|
|
|
| |
Note that this very slightly changes the existing API:
* removes unused refiner() getter
* adds copy assignment operator (before there was only an implicit copy constructor)
Change: 148933743
|
|
|
|
| |
Change: 148930597
|
|
|
|
| |
Change: 148930269
|
|
|
|
|
|
|
| |
This makes all such operations match the behavior of unsorted_segment_sum, for which completely missing segment IDs are output as zero initialized. In this case output is either zero or one initialized depending on the aggregation function. This makes such functions more easily usable for embeddings when some features are optional.
Zero/one initialization is only performed on the missing segment ids to avoid performance regressions.
Change: 148929039
|
|
|
|
|
| |
cuda-clang
Change: 148928329
|
|
|
|
| |
Change: 148926382
|
|
|
|
| |
Change: 148925596
|
|
|
|
| |
Change: 148922978
|
|
|
|
|
|
| |
tf.make_template(create_scope_now_=True) will now create its operations in the same name scope as its variables. This means that all variables and ops created by a template that is only applied once end up in the same scope.
Change: 148922089
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
values.
We used to rely on the mean being infinite, but oddly, Tensorboard gives us those errant values as strings:
[1, 256, 256, 0, 0, 0, 0, 0, "Infinity", "-Infinity", "NaN", "NaN"]
This breaks logic that checks for !== Infinity because a string is not Infinity. Something in our stack might have difficulty reading Infinity and NaN from JSON.
For now, I think we might want to resort to checking for valid values. This resolves a console error.
Change: 148920617
|
|
|
|
|
|
|
| |
storage.
Record persistent tensor memory and persistent memory (originally auxiliary memory) in the cost model.
Change: 148920117
|
|
|
|
| |
Change: 148919678
|
|
|
|
| |
Change: 148915145
|