Selects slices of `if_true` and `if_false` based on `condition`. This is a wrapper around tf.where() that allows it to select slices at any level of a tensor. (Recall that tf.where() can only select slices on the first axis or select individual elements). Example usage: condition = tf.
(condition, if_true, if_false)
| 48 | |
| 49 | |
| 50 | def where(condition, if_true, if_false): |
| 51 | """Selects slices of `if_true` and `if_false` based on `condition`. |
| 52 | |
| 53 | This is a wrapper around tf.where() that allows it to select slices at any |
| 54 | level of a tensor. (Recall that tf.where() can only select slices on the |
| 55 | first axis or select individual elements). |
| 56 | |
| 57 | Example usage: |
| 58 | |
| 59 | condition = tf.constant([[True, False], |
| 60 | [False, True]]) |
| 61 | if_true = tf.constant([[[1, 1], [2, 2]], |
| 62 | [[3, 3], [4, 4]]]) |
| 63 | if_false = tf.constant([[[5, 5], [6, 6]], |
| 64 | [[7, 7], [8, 8]]]) |
| 65 | |
| 66 | result = where(condition, if_true, if_false) |
| 67 | |
| 68 | assert result.eval() == [[[1, 1], [6, 6]], |
| 69 | [[7, 7], [4, 4]]] |
| 70 | |
| 71 | Args: |
| 72 | condition: <bool>[...] Condition for selecting slices. |
| 73 | if_true: <T>[condition.shape, ...] Values to use if `condition` is true. |
| 74 | if_false: <T>[condition.shape, ...] Values to use if `condition` is false. |
| 75 | |
| 76 | Returns: |
| 77 | <T>[if_true.shape] Values after conditional selection. |
| 78 | """ |
| 79 | # TODO(terrykoo): Rewrite when tf.where() supports broadcasting: b/78594182. |
| 80 | condition_rank = condition.shape.ndims |
| 81 | result_rank = if_true.shape.ndims |
| 82 | |
| 83 | # Use tf.where() directly if it can handle the args. |
| 84 | if condition_rank == 1 or condition_rank == result_rank: |
| 85 | return tf.where(condition, if_true, if_false) |
| 86 | |
| 87 | result_dims = shape(if_true) |
| 88 | inner_dims = result_dims[condition_rank:] |
| 89 | |
| 90 | # Flatten outer dims so we can select slices on axis=0. |
| 91 | condition = tf.reshape(condition, [-1]) |
| 92 | if_true = tf.reshape(if_true, [-1] + inner_dims) |
| 93 | if_false = tf.reshape(if_false, [-1] + inner_dims) |
| 94 | |
| 95 | result = tf.where(condition, if_true, if_false) |
| 96 | |
| 97 | # Restore original shape. |
| 98 | return tf.reshape(result, result_dims) |
| 99 | |
| 100 | |
| 101 | def padded_where(condition, length): |
nothing calls this directly
no test coverage detected
searching dependent graphs…