tf.linalg / det

TensorFlow 1 version

View source on GitHub

Returns the batched diagonal part of a batched tensor.

tf.linalg.diag_part(
    input, name='diag_part', k=0, padding_value=0,
    align='RIGHT_LEFT'
)

Used in the notebooks

Returns a tensor with the k[0]-th to k[1]-th diagonals of the batched input.

Assume input has r dimensions [I, J, ..., L, M, N]. Let max_diag_len be the maximum length among all diagonals to be extracted, max_diag_len = min(M + min(k[1], 0), N + min(-k[0], 0)) Let num_diags be the number of diagonals to extract, num_diags = k[1] - k[0] + 1.

If num_diags == 1, the output tensor is of rank r - 1 with shape [I, J, ..., L, max_diag_len] and values:

diagonal[i, j, ..., l, n]
  = input[i, j, ..., l, n+y, n+x] ; if 0 <= n+y < M and 0 <= n+x < N,
    padding_value                 ; otherwise.

where y = max(-k[1], 0), x = max(k[1], 0).

Otherwise, the output tensor has rank r with dimensions [I, J, ..., L, num_diags, max_diag_len] with values:

diagonal[i, j, ..., l, m, n]
  = input[i, j, ..., l, n+y, n+x] ; if 0 <= n+y < M and 0 <= n+x < N,
    padding_value                 ; otherwise.

where d = k[1] - m, y = max(-d, 0) - offset, and x = max(d, 0) - offset.

offset is zero except when the alignment of the diagonal is to the right.

offset = max_diag_len - diag_len(d) ; if (`align` in {RIGHT_LEFT, RIGHT_RIGHT}
                                           and `d >= 0`) or
                                         (`align` in {LEFT_RIGHT, RIGHT_RIGHT}
                                           and `d <= 0`)
         0                          ; otherwise

where diag_len(d) = min(cols - max(d, 0), rows + min(d, 0)).

The input must be at least a matrix.

For example:

input = np.array([[[1, 2, 3, 4],  # Input shape: (2, 3, 4)
                   [5, 6, 7, 8],
                   [9, 8, 7, 6]],
                  [[5, 4, 3, 2],
                   [1, 2, 3, 4],
                   [5, 6, 7, 8]]])

# A main diagonal from each batch.
tf.linalg.diag_part(input) ==> [[1, 6, 7],  # Output shape: (2, 3)
                                [5, 2, 7]]

# A superdiagonal from each batch.
tf.linalg.diag_part(input, k = 1)
  ==> [[2, 7, 6],  # Output shape: (2, 3)
       [4, 3, 8]]

# A band from each batch.
tf.linalg.diag_part(input, k = (-1, 2))
  ==> [[[3, 8, 0],  # Output shape: (2, 4, 3)
        [2, 7, 6],
        [1, 6, 7],
        [0, 5, 8]],
       [[3, 4, 0],
        [4, 3, 8],
        [5, 2, 7],
        [0, 1, 6]]]

# RIGHT_LEFT alignment.
tf.linalg.diag_part(input, k = (-1, 2), align="RIGHT_LEFT")
  ==> [[[0, 3, 8],  # Output shape: (2, 4, 3)
        [2, 7, 6],
        [1, 6, 7],
        [5, 8, 0]],
       [[0, 3, 4],
        [4, 3, 8],
        [5, 2, 7],
        [1, 6, 0]]]

# max_diag_len can be shorter than the main diagonal.
tf.linalg.diag_part(input, k = (-2, -1))
  ==> [[[5, 8],
        [0, 9]],
       [[1, 6],
        [0, 5]]]

# padding_value = 9
tf.linalg.diag_part(input, k = (1, 3), padding_value = 9)
  ==> [[[4, 9, 9],  # Output shape: (2, 3, 3)
        [3, 8, 9],
        [2, 7, 6]],
       [[2, 9, 9],
        [3, 4, 9],
        [4, 3, 8]]]