| 18f89c81d288f191abd56501ec6f86fe29265bdd |
|
17-Oct-2017 |
Eugene Brevdo <ebrevdo@google.com> |
[tf.contrib.seq2seq] Bugfixes to BeamSearchDecoder and GatherTree.
1. Begin the gather tree at the maximum sequence length across all beams (within the batch).
2. Take a second pass starting from t=0 and mask out any beam ids past the *first* beam occurrence of end_token.
3. Update the final sequence lengths to include the first <eos> token in the beam.
4. Update dynamic_decode to allow the BeamSearchDecoder to keep track of its own "finished" states, as the shuffling in the decoder confused the tracking mechanism in dynamic_decode. This fixes a bug where beam search decoding stops early.
5. Cap sequence length used in GatherTree to min(max_time, max_seq_len(b)) to avoid accessing memory outside the dimensions of input matrices.
Bugs caught by @bdaskalov on github and Pavel Sountsov. Proper solution and analysis thanks to Rui Zhao. Thanks all!
Fixes #13536.
PiperOrigin-RevId: 172471462
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 50b999a8336d19400ab75aea66fe46eca2f5fe0b |
|
28-Jun-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
Merge changes from github.
PiperOrigin-RevId: 160344052
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 891de62d0266113b067267ece48ea448665a5fbe |
|
27-Apr-2017 |
Eugene Brevdo <ebrevdo@google.com> |
[tf contrib seq2seq] Minor cleanups to beam search decoder.
Change: 154435114
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 17fdd80807b817b0940e572fea6e555c02ace71f |
|
20-Apr-2017 |
Eugene Brevdo <ebrevdo@google.com> |
[tf contrib seq2seq] Changes to dynamic decoding
1. dynamic_decode now returns a third parameter: the sequence lengths
from decoding (minibatch entries that finished earlier have a shorter
sequence length)
2. beam search decoder now uses the gather_tree C++ op
3. the gather_tree c++ op now expects sequence_length to be a matrix shaped
`[batch_size, beam_width]` (each beam may have its own sequence length).
Change: 153756869
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 72164f00a95eaa8468fe00e84fe356c753e2c164 |
|
20-Apr-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
BeamSearchDecoder:
1. Implemented in contrib/seq2seq.
2. Includes changes to attention_wrapper.py to accommodate the batch_size tiling.
3. Includes changes to decoder.py to accommodate having a finalize step.
Change: 153649516
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| ba7f9a78c8e3117fecbe6e2f84dcd0cf0e0dc8ef |
|
19-Apr-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
Improved seq2seq documentation.
Improved several docstrings.
Added an example in the guide.
Change: 153592069
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 916fcfb39a23afd96893bf85cb6f29c71a483642 |
|
05-Apr-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Fix dynamic_rnn transpose bug (can input/output non-3d tensors).
Also a few cleanups to RNN code.
Change: 152267628
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 03abac7f23e6e6864949b959435282729192692e |
|
17-Mar-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Updates to RNNCells to allow easy storage of attention TensorArray in the state.
The main change is that RNNCells that wrap other RNNCells now override self.zero_state to call the wrapped cell's zero_state and then (maybe) perform some post-processing... instead of relying on the state_size property to provide all information about the state.
Also made zero_state calls create ops inside their own name scope.
Change: 150413265
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| d46937f3684d3be2285f7401c2f471f542d147c2 |
|
14-Feb-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Remove complicated decoder API in favor of new object-oriented API in tf.contrib.seq2seq.
1. Expose the new classes and functions in tf.contrib.seq2seq
2. Delete old functions
3. Rename dynamic_decode_rnn -> dynamic_decode
Change: 147436900
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| fde892c4736f9383ba5e56324d6c6ec877bd7a0d |
|
03-Feb-2017 |
A. Unique TensorFlower <gardener@tensorflow.org> |
Change for internal compatibility.
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 3b168a56042ce5dcef4ce57db89e40c2ef6a839f |
|
31-Jan-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Refactor samplers in tf.contrib.seq2seq
1. Sampler now has two step calls: sample (which just return sample_ids),
and next_inputs (which accepts sample_ids, but also old state and returns
a possibly updated state). The updated state change is necessary for
TopK / beam search samplers.
2. Added a CustomSampler class that helps build samplers from simple helper fns.
In the near future, it'll be easy to add a ScheduledEmbeddingSampler and a standard EmbeddingSampler
(because all we need to provide is a new sampler method).
For now, everything's still part of the Sampler interface; but it's possible that
Sampler will go away in favor of BasicSamplingDecoder simply doing all the necessary work
and calling its own sample() and next_inputs() methods.
Change: 146051886
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 9a39d72cf1852c923455df17278358f6e1158e9d |
|
27-Jan-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Several cleanup changes to the new decoder in tf.contrib.seq2seq
1. Renamed the ArgmaxEmbeddingInferenceSampler to GreedyEmbeddingSampler
2. Moved max_time from this sampler to the decode_dynamic_rnn method
(it's more generally usable)
3. Added an impute_finished argument to decode_dynamic_rnn (default False)
which controls how states/outputs are handled for batch entries that are
finished.
Change: 145754947
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 8291536b277379ff37dbb8d3198d947776aaff2f |
|
25-Jan-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Add ArgmaxEmbeddingInferenceSampler to new tf.contrib.seq2seq API
Also add a properly caching varscope to the decoder fn.
Change: 145600420
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|
| 125fd7a5bb431d3da927e7a98182656c081a5e7c |
|
17-Jan-2017 |
Eugene Brevdo <ebrevdo@google.com> |
Initial design of new seq2seq Decoder class library.
Still hidden from tf.contrib.seq2seq namespace for rapid iteration.
Not for general use until the API has stabilized a bit.
Barebones documentation for now.
Change: 144733714
/external/tensorflow/tensorflow/contrib/seq2seq/python/ops/decoder.py
|