tft.deduplicate_tensor_per_row

Deduplicates each row (0-th dimension) of the provided tensor.

input_tensor A two-dimensional Tensor or SparseTensor. The first dimension is assumed to be the batch or "row" dimension, and deduplication is done on the 2nd dimension. If the Tensor is 1D it is returned as the equivalent SparseTensor since the "row" is a scalar can't be further deduplicated.
name Optional name for the operation.

A SparseTensor containing the unique set of values from each row of the input. Note: the original order of the input may not be preserved.