You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Return alignment history when decoding from an AttentionalRNNDecoder (requires TensorFlow 1.8+ when decoding with beam search)
Boolean parameter replace_unknown_target to replace unknown target tokens by the source token with the highest attention (requires a decoder that returns the alignment history)
Support for arbitrary transition layers in SequentialEncoder
Fixes and improvements
Fix sequence reduction when the maximum sequence length is not equal to the tensor time dimension (e.g. when splitting a batch for multi-GPU training)
The number of prefetched batches is automatically tuned when prefetch_buffer_size is not set (for TensorFlow 1.8+)