Returns a sequence mask, like tf.sequence_mask(). Unlike tf.sequence_mask(), this can also generate a transposed mask, which is convenient for working with time-major sequences. Args: lengths: [batch_size] Sequence lengths. maxlen: [] Maximum length, or None to compute the
(lengths, maxlen=None, dtype=tf.bool, transpose=False)
| 222 | |
| 223 | |
| 224 | def sequence_mask(lengths, maxlen=None, dtype=tf.bool, transpose=False): |
| 225 | """Returns a sequence mask, like tf.sequence_mask(). |
| 226 | |
| 227 | Unlike tf.sequence_mask(), this can also generate a transposed mask, which is |
| 228 | convenient for working with time-major sequences. |
| 229 | |
| 230 | Args: |
| 231 | lengths: <int>[batch_size] Sequence lengths. |
| 232 | maxlen: <int>[] Maximum length, or None to compute the max of `lengths`. |
| 233 | dtype: DType of the generated mask. |
| 234 | transpose: Whether to generate the transpose of the mask. |
| 235 | |
| 236 | Returns: |
| 237 | <dtype>[maxlen, batch_size] Sequence mask. |
| 238 | """ |
| 239 | with tf.name_scope("sequence_mask"): |
| 240 | if maxlen is None: |
| 241 | maxlen = tf.reduce_max(lengths) |
| 242 | positions = tf.range(maxlen, dtype=lengths.dtype) |
| 243 | |
| 244 | positions_singleton_axis = 1 if transpose else 0 |
| 245 | lengths_singleton_axis = 0 if transpose else 1 |
| 246 | positions = tf.expand_dims(positions, axis=positions_singleton_axis) |
| 247 | lengths = tf.expand_dims(lengths, axis=lengths_singleton_axis) |
| 248 | mask = positions < lengths |
| 249 | |
| 250 | if dtype != tf.bool: |
| 251 | mask = tf.cast(mask, dtype) |
| 252 | return mask |
nothing calls this directly
no outgoing calls
no test coverage detected
searching dependent graphs…