| Commit message (Collapse) | Author | Age |
|
|
|
| |
PiperOrigin-RevId: 209679086
|
|
|
|
| |
PiperOrigin-RevId: 184347081
|
|
|
|
|
|
|
| |
I'm trying to make core/kernels independent of protos. Currently the dtype ResourceHandle is itself a proto. After this CL, ResourceHandle is a normal C++ type which gets converted to/from ResourceHandleProto at (de)serialization time.
RELNOTES: n/a
PiperOrigin-RevId: 160329002
|
|
|
|
| |
Change: 142707321
|
|
|
|
| |
Change: 137329621
|
|
|
|
| |
Change: 134714467
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Currently this does not support broadcasts, so the r-value's shape must
match the shape of the slice.
- Only assignments to variables are supported.
e.g. of usage
op = foo[3:5].assign(tf.constant([1,2]))
sess.run(op)
Change: 131837883
|
|
|
|
| |
Change: 127951209
|
|
|
|
|
|
|
|
|
| |
StridedSliceGrad op implements the gradient of StridedSlice.
Also implement python benchmark for StridedSlice and simple Slice.
Fix bugs in special case optimizations in StridedSlice.
(Toward resolving bug #206)
Change: 125789921
|
|
This extends slicing to allow a more complete set of the basic slicing
in NumPy. For example, you can now do negative strides while zero strides
are still disallowed.
NOTE: This change does not yet hook up the new slicing behavior to the
python getitem operator, because we do not yet support gradients
(unlike the simpler slice we are extending).
NOTE: This change does not introduce slicing using tensors as arguments
(advanced indexing in NumPy).
foo[5:1:-2]
You can also do ellipsis
foo[5:1:-2,...,3:4]
In addition you can add new dimensions
foo[5:, tf.newaxis, ...]
Fixes part of #206.
Change: 124998811
|