| Commit message (Collapse) | Author | Age |
... | |
|
|
|
|
|
|
|
| |
tensor is intialized early (because we ran out of memory) and report
it as such.
Fixes #7025.
Change: 154603030
|
|
|
|
|
|
| |
You can use a tuple to label a point along an axis, so tuples should not be
converted into lists of points for indexing.
Change: 154603014
|
|
|
|
| |
Change: 154602604
|
|
|
|
| |
Change: 154598657
|
|
|
|
|
|
|
|
|
|
|
| |
Several nn_ops were registered for ints accidentally. Some of these even have
kernel registrations for ints which do very strange things. The ops without
kernels registered have been restricted to floating point types, and warnings
have been added for those with kernels. Other ops were registered with more
float types than they had kernels for.
Fixes #9317. Fixes #5539.
Change: 154595629
|
|
|
|
|
| |
Fixes #9441.
Change: 154593819
|
|
|
|
|
|
| |
are Variables.
Change: 154593246
|
|
|
|
|
| |
Add comparison of ordering of HLO instructions which are in different computations using the call graph. Previously, instructions in different computations were considered unordered. Ordering these instructions improves buffer liveness analysis and may enable better buffer sharing between values in different computations.
Change: 154592912
|
|
|
|
|
|
|
| |
sparse_tensor_dense_matmul.
Fixes #8131
Change: 154591681
|
|
|
|
| |
Change: 154591046
|
|
|
|
|
| |
Fixes #5761
Change: 154590417
|
|
|
|
| |
Change: 154590336
|
|
|
|
| |
Change: 154589265
|
|
|
|
| |
Change: 154589003
|
|
|
|
|
|
|
| |
Removes incorrect "per batch" language.
Fixes #3824.
Change: 154587896
|
|
|
|
|
|
| |
ops doesn't share the scope.
Change: 154586285
|
|
|
|
| |
Change: 154586226
|
|
|
|
| |
Change: 154585848
|
|
|
|
| |
Change: 154584953
|
|
|
|
| |
Change: 154584844
|
|
|
|
| |
Change: 154584479
|
|
|
|
| |
Change: 154583581
|
|
|
|
|
|
| |
check. The XLA GPU JIT can only be used if the Tensorflow GPU device exists, which requires that we compile with CUDA.
Change: 154582511
|
|
|
|
| |
Change: 154582184
|
|
|
|
|
|
|
| |
The new version of the shape documentation for both ops is hopefully correct and doesn't introduce temporary variables. Additionally, tf.scatter_nd warns about nondeterministic behavior if indices contains duplicates.
Fixes #8245. Fixes #8102.
Change: 154581230
|
|
|
|
| |
Change: 154579657
|
|
|
|
|
|
| |
the performance by almost 20% on my GPU (from 55 images/s to 62 images/s)
Change: 154578655
|
|
|
|
| |
Change: 154578604
|
|
|
|
| |
Change: 154577958
|
|
|
|
| |
Change: 154575206
|
|
|
|
| |
Change: 154572855
|
|
|
|
|
|
|
|
|
| |
constant folder will constant-fold DT_INT32 operators.
Tensorflow core: Add support for constant folding of nodes with control dependencies.
Mark candidate sampling ops as stateful; the constant-folder was now constant-folding them where it should not.
Change: 154572578
|
|
|
|
| |
Change: 154571409
|
|
|
|
| |
Change: 154569724
|
|
|
|
| |
Change: 154568808
|
|
|
|
| |
Change: 154567986
|
|
|
|
| |
Change: 154566292
|
|
|
|
| |
Change: 154565789
|
|
|
|
| |
Change: 154564406
|
|
|
|
| |
Change: 154562449
|
|
|
|
| |
Change: 154556754
|
|
|
|
|
| |
Fixes #8809
Change: 154553141
|
|
|
|
| |
Change: 154548095
|
|
|
|
| |
Change: 154547403
|
|
|
|
| |
Change: 154507631
|
|
|
|
| |
Change: 154501275
|
|
|
|
|
| |
"reduction_ops_all.cc",
Change: 154499070
|
|
|
|
|
|
| |
align_corners=True. In particular, we change floorf() to roundf() in discretizing the continuous coordinates when align_corners=True.
Change: 154492447
|
|
|
|
|
| |
cl/154253719 added a CHECK which fails if a value (LogicalBuffer) is used more than once by an instruction. This resulted in duplicate values in a list of uses.
Change: 154485725
|
|
|
|
| |
Change: 154483923
|