LogSoftmax#

  • Domain: ai.onnx

  • Since version: 13

The operator computes the log of softmax values for the given input:

LogSoftmax(input, axis) = Log(Softmax(input, axis=axis))

The “axis” attribute indicates the dimension along which LogSoftmax will be performed. The output tensor has the same shape as the input tensor.

Inputs

  • input (T): The input tensor of rank >= axis.

Outputs

  • output (T): The output values with the same shape as the input tensor.

Attributes

  • axis (int): Describes the dimension Softmax will be performed on. Negative value means counting dimensions from the back.

Type Constraints

  • T: Constrain input and output types to float tensors. Allowed types: tensor(bfloat16), tensor(double), tensor(float), tensor(float16).

Examples#

test_cc_logsoftmax_axis_0

Node:
  LogSoftmax(input) -> (output)
  Attributes:
    axis = 0

Inputs:
  input: shape=(2, 3), dtype=float32
    [[ 1.,  2.,  3.],
     [ 4.,  1., -1.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-3.0485873 , -0.31326175, -0.01814985],
     [-0.04858732, -1.3132617 , -4.01815   ]]

test_cc_logsoftmax_axis_1

Node:
  LogSoftmax(input) -> (output)
  Attributes:
    axis = 1

Inputs:
  input: shape=(2, 3), dtype=float32
    [[ 1.,  2.,  3.],
     [ 4.,  1., -1.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-2.407606  , -1.4076059 , -0.4076059 ],
     [-0.05498505, -3.054985  , -5.054985  ]]

test_cc_logsoftmax_axis_2

Node:
  LogSoftmax(input) -> (output)
  Attributes:
    axis = 2

Inputs:
  input: shape=(2, 2, 3), dtype=float32
    [[[ 1. ,  2. ,  3. ],
      [ 4. ,  1. , -1. ]],

     [[ 0.5, -0.5,  2. ],
      [-2. ,  1.5,  0. ]]]

Outputs:
  output: shape=(2, 2, 3), dtype=float32
    [[[-2.407606  , -1.4076059 , -0.4076059 ],
      [-0.05498505, -3.054985  , -5.054985  ]],

     [[-1.7663679 , -2.766368  , -0.2663679 ],
      [-3.725802  , -0.22580206, -1.7258021 ]]]

test_cc_logsoftmax_default_axis

Node:
  LogSoftmax(input) -> (output)

Inputs:
  input: shape=(2, 3), dtype=float32
    [[ 1.,  2.,  3.],
     [ 4.,  1., -1.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-2.407606  , -1.4076059 , -0.4076059 ],
     [-0.05498505, -3.054985  , -5.054985  ]]

test_cc_logsoftmax_example_1

Node:
  LogSoftmax(input) -> (output)
  Attributes:
    axis = 1

Inputs:
  input: shape=(2, 3), dtype=float32
    [[1., 2., 3.],
     [1., 2., 3.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-2.407606 , -1.4076059, -0.4076059],
     [-2.407606 , -1.4076059, -0.4076059]]

test_cc_logsoftmax_large_number

Node:
  LogSoftmax(input) -> (output)

Inputs:
  input: shape=(2, 3), dtype=float32
    [[1000., 1001., 1002.],
     [1002., 1001., 1000.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-2.4075928 , -1.4075928 , -0.40759277],
     [-0.40759277, -1.4075928 , -2.4075928 ]]

test_cc_logsoftmax_negative_axis

Node:
  LogSoftmax(input) -> (output)
  Attributes:
    axis = -1

Inputs:
  input: shape=(2, 3), dtype=float32
    [[ 1.,  2.,  3.],
     [ 4.,  1., -1.]]

Outputs:
  output: shape=(2, 3), dtype=float32
    [[-2.407606  , -1.4076059 , -0.4076059 ],
     [-0.05498505, -3.054985  , -5.054985  ]]

Differences with previous version (11)#

SchemaDiff: LogSoftmax (domain 'ai.onnx')

  • old version: 11

  • new version: 13

  • breaking: yes

Breaking reasons:

  • attribute ‘axis’ (changed): default value changed 1 -> -1

Attributes:

  • [BREAKING] changed ‘axis’: default value changed 1 -> -1

Type constraints:

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

Documentation:

  • line similarity: 0.18 (+6/-3 lines)

--- LogSoftmax v11
+++ LogSoftmax v13
@@ -1,4 +1,7 @@

-The operator computes the logsoftmax (log of softmax) values for the given input.
-The "axis" attribute indicates the dimension along which LogSoftmax is
-performed. The output tensor has the same shape as the input tensor.
+The operator computes the log of softmax values for the given input:
+
+ LogSoftmax(input, axis) = Log(Softmax(input, axis=axis))
+
+The "axis" attribute indicates the dimension along which LogSoftmax
+will be performed. The output tensor has the same shape as the input tensor.

Version History#