tf.raw_ops.LoadAndRemapMatrix

Loads a 2-D (matrix) Tensor with name old_tensor_name from the checkpoint

tf.raw_ops.LoadAndRemapMatrix(
    ckpt_path, old_tensor_name, row_remapping, col_remapping, initializing_values,
    num_rows, num_cols, max_rows_in_memory=-1, name=None
)

at ckpt_path and potentially reorders its rows and columns using the specified remappings.

Most users should use one of the wrapper initializers (such as tf.contrib.framework.load_and_remap_matrix_initializer) instead of this function directly.

The remappings are 1-D tensors with the following properties:

• row_remapping must have exactly num_rows entries. Row i of the output matrix will be initialized from the row corresponding to index row_remapping[i] in the old Tensor from the checkpoint.
• col_remapping must have either 0 entries (indicating that no column reordering is needed) or num_cols entries. If specified, column j of the output matrix will be initialized from the column corresponding to index col_remapping[j] in the old Tensor from the checkpoint.
• A value of -1 in either of the remappings signifies a "missing" entry. In that case, values from the initializing_values tensor will be used to fill that missing row or column. If row_remapping has r missing entries and col_remapping has c missing entries, then the following condition must be true:

(r * num_cols) + (c * num_rows) - (r * c) == len(initializing_values)

The remapping tensors can be generated using the GenerateVocabRemapping op.

As an example, with row_remapping = [1, 0, -1], col_remapping = [0, 2, -1], initializing_values = [0.5, -0.5, 0.25, -0.25, 42], and w(i, j) representing the value from row i, column j of the old tensor in the checkpoint, the output matrix will look like the following:

[[w(1, 0), w(1, 2), 0.5], [w(0, 0), w(0, 2), -0.5], [0.25, -0.25, 42]]