| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 214700693
|
| |
|
|\ |
|
| |
| |
| |
| | |
PiperOrigin-RevId: 207615481
|
|/ |
|
|
|
|
|
|
|
|
| |
* implement tf.arg, closes #483
* Remove GPU support for arg op
* rename arg to angle
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Implemented asinh op
* Implemented acosh op
* Fixed test functions for acosh and asinh
* Implemented atanh op
* Fixed test cases for atanh
|
|
|
|
|
|
|
|
|
|
| |
* Implemented sinh and cosh #7531
* Removed Eigen::half from cosh and sinh definitions
Refer to Issue #7531 and Pull Request #7628
* Fixed the gradient for sinh in math_grad_test.cc
|
|
|
|
|
|
|
|
| |
* TF-549 Adds unsorted segment max Op
* Cosmetic change
* Add todo comment
|
|
|
|
|
| |
of Define() where that doesn't work, switch to Create() instead.
Change: 142718606
|
|
|
|
| |
Change: 142198004
|
|
|
|
| |
Change: 140917655
|
|
|
|
| |
Change: 139958241
|
|
|
|
| |
Change: 139240711
|
|
|
|
| |
Change: 138787117
|
|
|
|
| |
Change: 131294380
|
|
|
|
|
|
|
|
| |
d(x ^ y)/dy mentions log(x), which creates false singularities if x = y.
Instead, use x > 0 ? log(x) : 0.
Fixes #2295.
Change: 127762504
|
|
|
|
| |
Change: 124012080
|
|
|
|
| |
Change: 123900938
|
|
|
|
| |
Change: 123026122
|
|
|
|
|
|
| |
CUDA 7.5 or higher. (If not, the GPU tests will also not be run,
so that the tests as a whole will keep passing on CUDA 7.0.)
Change: 122566230
|
|
|
|
| |
Change: 121382365
|
|
|
|
|
|
|
|
|
| |
This also includes updating parts of the Python test framework to handle fp16.
fp16 is too inaccurate to do numerical gradients unless a lot of care is taken,
so for fp16, we compare fp16 theoretical gradients to fp32 numerical ones.
This means that the gradient check doesn't also implicitly test the function
itself, so we will need to rely on the numpy tests for those.
Change: 119948035
|
|
|
|
|
|
|
|
| |
until needed the backprop-ed "dy" is ready.
E.g., y = tanh(x). dx = dy * (1 - y^2). We don't need to compute y^2
until dy is computed.
Change: 114781664
|
|
|
|
| |
Change: 112920860
|
|
|
|
|
| |
After this we can replace port.h with types.h.
Change: 112727463
|
|
|
|
|
|
|
| |
The implementation divides the weighted sum by sqrt(n) or divides by the sqrt(sum(x^2)) if sp_weights are specified on embedding_lookup_sparse.
Also implements math_ops.sparse_segment_sqrtn and its corresponding math_ops.sparse_segment_sqrtn_grad.
Change: 111889513
|
|
since they aren't part of the generated API. remove cc_library
in py_library which breaks OSS build.
Change: 111542333
|