tf.linalg / normalize

TensorFlow 1 version

View source on GitHub

Returns a batched matrix tensor with new batched diagonal values.

tf.linalg.set_diag(
    input, diagonal, name='set_diag', k=0, align='RIGHT_LEFT'
)

Given input and diagonal, this operation returns a tensor with the same shape and values as input, except for the specified diagonals of the innermost matrices. These will be overwritten by the values in diagonal.

input has r+1 dimensions [I, J, ..., L, M, N]. When k is scalar or k[0] == k[1], diagonal has r dimensions [I, J, ..., L, max_diag_len]. Otherwise, it has r+1 dimensions [I, J, ..., L, num_diags, max_diag_len]. num_diags is the number of diagonals, num_diags = k[1] - k[0] + 1. max_diag_len is the longest diagonal in the range [k[0], k[1]], max_diag_len = min(M + min(k[1], 0), N + min(-k[0], 0))

The output is a tensor of rank k+1 with dimensions [I, J, ..., L, M, N]. If k is scalar or k[0] == k[1]:

output[i, j, ..., l, m, n]
  = diagonal[i, j, ..., l, n-max(k[1], 0)] ; if n - m == k[1]
    input[i, j, ..., l, m, n]              ; otherwise

Otherwise,

output[i, j, ..., l, m, n]
  = diagonal[i, j, ..., l, diag_index, index_in_diag] ; if k[0] <= d <= k[1]
    input[i, j, ..., l, m, n]                         ; otherwise

where d = n - m, diag_index = k[1] - d, and index_in_diag = n - 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)).

For example:

# The main diagonal.
input = np.array([[[7, 7, 7, 7],              # Input shape: (2, 3, 4)
                   [7, 7, 7, 7],
                   [7, 7, 7, 7]],
                  [[7, 7, 7, 7],
                   [7, 7, 7, 7],
                   [7, 7, 7, 7]]])
diagonal = np.array([[1, 2, 3],               # Diagonal shape: (2, 3)
                     [4, 5, 6]])
tf.matrix_set_diag(input, diagonal)
  ==> [[[1, 7, 7, 7],  # Output shape: (2, 3, 4)
        [7, 2, 7, 7],
        [7, 7, 3, 7]],
       [[4, 7, 7, 7],
        [7, 5, 7, 7],
        [7, 7, 6, 7]]]

# A superdiagonal (per batch).
tf.matrix_set_diag(input, diagonal, k = 1)
  ==> [[[7, 1, 7, 7],  # Output shape: (2, 3, 4)
        [7, 7, 2, 7],
        [7, 7, 7, 3]],
       [[7, 4, 7, 7],
        [7, 7, 5, 7],
        [7, 7, 7, 6]]]

# A band of diagonals.
diagonals = np.array([[[9, 1, 0],  # Diagonal shape: (2, 4, 3)
                       [6, 5, 8],
                       [1, 2, 3],
                       [0, 4, 5]],
                      [[1, 2, 0],
                       [5, 6, 4],
                       [6, 1, 2],
                       [0, 3, 4]]])
tf.matrix_set_diag(input, diagonals, k = (-1, 2))
  ==> [[[1, 6, 9, 7],  # Output shape: (2, 3, 4)
        [4, 2, 5, 1],
        [7, 5, 3, 8]],
       [[6, 5, 1, 7],
        [3, 1, 6, 2],
        [7, 4, 2, 4]]]

# RIGHT_LEFT alignment.
diagonals = np.array([[[0, 9, 1],  # Diagonal shape: (2, 4, 3)
                       [6, 5, 8],
                       [1, 2, 3],
                       [4, 5, 0]],
                      [[0, 1, 2],
                       [5, 6, 4],
                       [6, 1, 2],
                       [3, 4, 0]]])
tf.matrix_set_diag(input, diagonals, k = (-1, 2), align="RIGHT_LEFT")
  ==> [[[1, 6, 9, 7],  # Output shape: (2, 3, 4)
        [4, 2, 5, 1],
        [7, 5, 3, 8]],
       [[6, 5, 1, 7],
        [3, 1, 6, 2],
        [7, 4, 2, 4]]]