diff options
author | 2017-08-22 11:10:29 -0700 | |
---|---|---|
committer | 2017-08-22 11:14:26 -0700 | |
commit | cb3314159fe102419289d394246d7ac9c2a422c1 (patch) | |
tree | 71548916d8b309f63935fb29bd846df2e99d87be /RELEASE.md | |
parent | 570a922d2b7e8dd33f2ffb4988ea3f306d34a192 (diff) |
Bugfix to DropoutWrapper: never drop out the "c" (memory) of an LSTMStateTuple.
This is not exactly a backwards compatible change, in that it modifies the
behavior of DropoutWrapper in its most typical use case; but it is a bugfix:
Dropout and variational dropout should now correctly train with LSTM.
Fixes #11650
PiperOrigin-RevId: 166085514
Diffstat (limited to 'RELEASE.md')
-rw-r--r-- | RELEASE.md | 12 |
1 files changed, 12 insertions, 0 deletions
diff --git a/RELEASE.md b/RELEASE.md index 3203f0aec1..ec66e555ad 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -1,3 +1,15 @@ +# Release 1.4.0 + +## Major Features And Improvements + +## Bug Fixes and Other Changes +* `tf.nn.rnn_cell.DropoutWrapper` is now more careful about dropping out LSTM + states. Specifically, it no longer ever drops the `c` (memory) state of an + `LSTMStateTuple`. The new behavior leads to proper dropout behavior + for LSTMs and stacked LSTMs. This bug fix follows recommendations from + published literature, but is a behavioral change. State dropout behavior + may be customized via the new `dropout_state_filter_visitor` argument. + # Release 1.3.0 See also [TensorBoard 0.1.4](https://github.com/tensorflow/tensorboard/releases/tag/0.1.4) release notes. |