Interface | Description |
---|---|
Accumulation |
An accumulation is an op that given:
x -> the origin ndarray y -> the pairwise ndarray n -> the number of times to accumulate Of note here in the extra arguments. |
BroadcastOp |
A broad cast op is one where a scalar
or less rank array
is broadcast to fill
a bigg er array.
|
CustomOp |
This interface describe "custom operations.
|
GradientOp |
A gradient op represents
a jacobian operation
|
GridOp |
MetaOp is special op, that contains multiple ops
|
IndexAccumulation |
An index accumulation is an operation that returns an index within
a NDArray.
Examples of IndexAccumulation operations include finding the index of the minimim value, index of the maximum value, index of the first element equal to value y, index of the maximum pair-wise difference between two NDArrays X and Y etc. Index accumulation is similar to Accumulation in that both are
accumulation/reduction operations, however index accumulation returns
an integer corresponding to an index, rather than a real (or complex)
value.Index accumulation operations generally have 3 inputs: x -> the origin ndarray y -> the pairwise ndarray (frequently null/not applicable) n -> the number of times to accumulate Note that IndexAccumulation op implementations should be stateless (other than the final result and x/y/n arguments) and hence threadsafe, such that they may be parallelized using the update, combineSubResults and set/getFinalResults methods. |
LossFunction |
A loss function for computing
the delta between two arrays
|
MetaOp |
MetaOp is special op, that contains multiple ops
|
Module |
A Module is a
CustomOp
with varying input arguments
and automatically calculated outputs
defined at a higher level than c++. |
Op |
An op is defined as follows:
opName: opName of the operation
x: the origin ndarray
y: the ndarray to parse in parallel
z: the resulting buffer
n: the number of elements to iterate over
where x is the origin ndarray,
y, is a pairwise op
over n elements in the ndarray
stored in result z
This is followed from the standard template for a BLAS operation
such that given a linear buffer, a function defines 3 buffers (x,y,z)
and the associated strides and offsets (handled by the ndarrays in this case)
|
RandomOp | |
ScalarOp |
Applies a scalar
along a bigger input array.
|
TransformOp |
Transform operation:
stores the result in an ndarray
|
Class | Description |
---|---|
BaseAccumulation |
Base class for accumulation, initiates the initial entry
with respect to the child class.
|
BaseBroadcastOp | |
BaseGradientOp |
A gradient op always makes the following assumptions:
there is always a y (beacuse of backpropagating
or using the chain rule)
and that it is special exec (for now)
This op opType sis meant to be used
to build derivative operations.
|
BaseIndexAccumulation |
Index based reduction algo
|
BaseModule |
Abstract base class for
Module
that handles Dynamic ops and handles nesting. |
BaseOp |
Base op.
|
BaseScalarOp |
Base scalar operation
|
BaseTransformOp |
A base op for basic getters and setters
|
CustomOpDescriptor |
This class is simple POJO that contains basic information about CustomOp
|
DefaultOpConverter | |
DynamicCustomOp |
Basic implementation for CustomOp
|
DynamicCustomOp.DynamicCustomOpsBuilder | |
DynamicCustomOp.SameDiffBuilder | |
NoOp | |
ShapeOp |
Shape manipulation ops
|
Enum | Description |
---|---|
Op.Type |
Copyright © 2018. All rights reserved.