ScatterElements - version 13#

This page documents version 13 of operator ScatterElements. See ScatterElements for the latest version (since version 18).

  • Domain: ai.onnx

  • Since version: 13

ScatterElements takes three inputs data, updates, and indices of the same rank r >= 1 and an optional attribute axis that identifies an axis of data (by default, the outer-most axis, that is axis 0). The output of the operation is produced by creating a copy of the input data, and then updating its value to values specified by updates at specific index positions specified by indices. Its output shape is the same as the shape of data.

For each entry in updates, the target index in data is obtained by combining the corresponding entry in indices with the index of the entry itself: the index-value for dimension = axis is obtained from the value of the corresponding entry in indices and the index-value for dimension != axis is obtained from the index of the entry itself.

reduction allows specification of an optional reduction operation, which is applied to all values in updates tensor into output at the specified indices. In cases where reduction is set to “none”, indices should not have duplicate entries: that is, if idx1 != idx2, then indices[idx1] != indices[idx2]. For instance, in a 2-D tensor case, the update corresponding to the [i][j] entry is performed as below:

output[indices[i][j]][j] = updates[i][j] if axis = 0,
output[i][indices[i][j]] = updates[i][j] if axis = 1,

When reduction is set to “add”, the update corresponding to the [i][j] entry is performed as below:

output[indices[i][j]][j] += updates[i][j] if axis = 0,
output[i][indices[i][j]] += updates[i][j] if axis = 1,

When reduction is set to “mul”, the update corresponding to the [i][j] entry is performed as below:

output[indices[i][j]][j] *= updates[i][j] if axis = 0,
output[i][indices[i][j]] *= updates[i][j] if axis = 1,

This operator is the inverse of GatherElements. It is similar to Torch’s Scatter operation.

Example 1:

data = [
    [0.0, 0.0, 0.0],
    [0.0, 0.0, 0.0],
    [0.0, 0.0, 0.0],
]
indices = [
    [1, 0, 2],
    [0, 2, 1],
]
updates = [
    [1.0, 1.1, 1.2],
    [2.0, 2.1, 2.2],
]
output = [
    [2.0, 1.1, 0.0]
    [1.0, 0.0, 2.2]
    [0.0, 2.1, 1.2]
]

Example 2:

data = [[1.0, 2.0, 3.0, 4.0, 5.0]]
indices = [[1, 3]]
updates = [[1.1, 2.1]]
axis = 1
output = [[1.0, 1.1, 3.0, 2.1, 5.0]]

Inputs

  • data (T): Tensor of rank r >= 1.

  • indices (Tind): Tensor of int32/int64 indices, of r >= 1 (same rank as input). All index values are expected to be within bounds [-s, s-1] along axis of size s. It is an error if any of the index values are out of bounds.

  • updates (T): Tensor of rank r >=1 (same rank and shape as indices)

Outputs

  • output (T): Tensor of rank r >= 1 (same rank as input).

Attributes

  • axis (int): Which axis to scatter on. Negative value means counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(data).

Type Constraints

  • T: Input and output types can be of any tensor type. Allowed types: tensor(bfloat16), tensor(bool), tensor(complex128), tensor(complex64), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(string), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8).

  • Tind: Constrain indices to integer types Allowed types: tensor(int32), tensor(int64).

Differences with previous version (11)#

SchemaDiff: ScatterElements (domain 'ai.onnx')

  • old version: 11

  • new version: 13

  • breaking: no

Type constraints:

  • changed ‘T’: added types: [‘tensor(bfloat16)’]

Documentation:

  • line similarity: 0.43 (+40/-27 lines)

--- ScatterElements v11
+++ ScatterElements v13
@@ -12,41 +12,54 @@
 entry in `indices` and the index-value for dimension != axis is obtained from the
 index of the entry itself.

-For instance, in a 2-D tensor case, the update corresponding to the [i][j] entry
-is performed as below:
+`reduction` allows specification of an optional reduction operation, which is applied to all values in `updates`
+tensor into `output` at the specified `indices`.
+In cases where `reduction` is set to "none", indices should not have duplicate entries: that is, if idx1 != idx2,
+then indices[idx1] != indices[idx2]. For instance, in a 2-D tensor case, the update
+corresponding to the [i][j] entry is performed as below:
 ```
-  output[indices[i][j]][j] = updates[i][j] if axis = 0,
-  output[i][indices[i][j]] = updates[i][j] if axis = 1,
+output[indices[i][j]][j] = updates[i][j] if axis = 0,
+output[i][indices[i][j]] = updates[i][j] if axis = 1,
+```
+When `reduction` is set to "add", the update corresponding to the [i][j] entry is performed as below:
+```
+output[indices[i][j]][j] += updates[i][j] if axis = 0,
+output[i][indices[i][j]] += updates[i][j] if axis = 1,
+```
+When `reduction` is set to "mul", the update corresponding to the [i][j] entry is performed as below:
+```
+output[indices[i][j]][j] *= updates[i][j] if axis = 0,
+output[i][indices[i][j]] *= updates[i][j] if axis = 1,
 ```

 This operator is the inverse of GatherElements. It is similar to Torch's Scatter operation.

 Example 1:
 ```
-  data = [
-      [0.0, 0.0, 0.0],
-      [0.0, 0.0, 0.0],
-      [0.0, 0.0, 0.0],
-  ]
-  indices = [
-      [1, 0, 2],
-      [0, 2, 1],
-  ]
-  updates = [
-      [1.0, 1.1, 1.2],
-      [2.0, 2.1, 2.2],
-  ]
-  output = [
-      [2.0, 1.1, 0.0]
-      [1.0, 0.0, 2.2]
-      [0.0, 2.1, 1.2]
-  ]
+data = [
+    [0.0, 0.0, 0.0],
+    [0.0, 0.0, 0.0],
+    [0.0, 0.0, 0.0],
+]
+indices = [
+    [1, 0, 2],
+    [0, 2, 1],
+]
+updates = [
+    [1.0, 1.1, 1.2],
+    [2.0, 2.1, 2.2],
+]
+output = [
+    [2.0, 1.1, 0.0]
+    [1.0, 0.0, 2.2]
+    [0.0, 2.1, 1.2]
+]
 ```
 Example 2:
 ```
-  data = [[1.0, 2.0, 3.0, 4.0, 5.0]]
-  indices = [[1, 3]]
-  updates = [[1.1, 2.1]]
-  axis = 1
-  output = [[1.0, 1.1, 3.0, 2.1, 5.0]]
+data = [[1.0, 2.0, 3.0, 4.0, 5.0]]
+indices = [[1, 3]]
+updates = [[1.1, 2.1]]
+axis = 1
+output = [[1.0, 1.1, 3.0, 2.1, 5.0]]
 ```