| Collect the last N rows from input data. The
| purpose is to keep track of data accross batches,
| so for example suppose the LastNWindowCollector is
| called successively with the following input data
|
| [1, 2, 3, 4]
| [5, 6, 7]
| [8, 9, 10, 11]
|
| And the number of items is set to 6, then the
| output after the 3rd call will contain the
| following elements:
|
| [6, 7, 8, 9, 10, 11]
|
| No guarantee is made on the ordering of elements
| in input.
|
| So a valid value for output could have been
|
| [11, 10, 9, 8, 7, 6]
|
| Also, this method works for any order tensor,
| treating the first dimension as input rows and
| keeping the last N rows seen as input. So for
| instance:
|
| [[1, 2], [2, 3], [3, 4], [4, 5]]
| [[5, 6], [6, 7], [7, 8]]
| [[8, 9], [9, 10], [10, 11], [11, 12]]
|
| A possible output would be
|
| [[6, 7], [7, 8], [8, 9], [9, 10], [10, 11], [11, 12]]
|
| This is not thread safe unless a mutex is given.