| Commit message (Collapse) | Author | Age |
|
|
|
| |
Change: 139959769
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
BatchCholesky
BatchCholeskyGrad
BatchMatrixDeterminant
BatchMatrixInverse
BatchMatrixSolve
BatchMatrixSolveLs
BatchMatrixTriangularSolve
BatchSelfAdjointEig
BatchSelfAdjointEigV2
BatchSvd
At this point, the non-batch versions were identical to the batch versions except for a check during shape inference restricting the matrix inputs to be rank 2.
NOTICE: This is a non-backwards compatible API change.
Change: 132692980
|
|
|
|
|
|
|
|
|
| |
ParseSingleSequenceExample.
The attribute parsing code was moved over verbatim to example_proto_helper.*.
Change TF_CHECK_OK to TF_ASSERT_OK in all shape inference tests under core.
Change: 130041336
|
|
|
|
|
|
|
|
|
|
|
| |
as separate outputs. This is implemented as a new set of kernels SelfAdjointEigV2 and BatchSelfAdjointEigV2. {batch_}self_adjoint_eig become Python wrappers that call the new kernels.
Adds new ops self_adjoint_eigvals and batch_self_adjoint_eigvals that compute the eigenvalues but not the eigenvectors, which is faster and uses less memory than computing both.
NOTICE: This changes the public API and TensorFlow models calling tf.self_adjoint_eig or tf.batch_self_adjoint_eig from Python will most likely break. Models saved as serialized graphs should be unaffected.
This CL also cleans up the python shape inference functions in linalg_ops.py a bit.
Change: 129390113
|
|
|
|
|
|
|
|
| |
This also adds Min(), Max(), and Subtract() operators and a few convenience methods to the InferenceContext. Change test utils to emit a human readable error message in case the user forgot to set the inference function.
Refactored shape_inference* a bit to enforce the invariant that a Dimension or DimensionOrConstant is always non-negative or equal to InferenceContext::kUnknownDim.
This made it possible to tighten & simplify the arithmetic operations a bit.
Change: 129385995
|
|
|
|
|
|
|
|
|
| |
- In shape inference test constructor, record an error and check it later in
testing - this makes for nicer errors when there is a bad input spec.
- Check that the number of inputs passed in the input spec matches what the
NodeDef declares; fix up violations.
- Remove constexpr kUnknownDim static in files.
Change: 128188634
|
|
Also changed python inference function for MatrixSolve and MatrixSolveLs to use
with_rank instead of with_rank_at_least for rhs.
Change: 127781866
|