Highlighted Changes

• This release includes code for working with the DROP dataset, including the official evaluation script, a DatasetReader, and the NAQANet model. (#2559, #2556 and #2560)
• We added a no-op trainer that allows you to create AllenNLP model archives for programmatic baselines, alternatively trained models, etc. (#2610)

Breaking Changes

1. In #2607 we changed the (default) SpacyWordSplitter to return allennlp Tokens (which are compact, efficient NamedTuples) rather than spacy.tokens.Tokens. This was done primarily to decrease memory usage for large datasets; secondarily to play nicer with the multiprocess dataset reader.

This is a breaking change in two ways, neither of which should affect most users:

• in theory everyone should be programming to the Token abstraction that's shared between both implementations, but it's possible that someone could be relying on having the actual spacy token, in which case they would need need to configure their word splitter with keep_spacy_tokens=True.

• a NamedTuple can't have different constructor parameters and field names. Our previous Token implementation used e.g. pos as the name of the constructor argument but then pos_ as the name of the field. Converting this to a namedtuple meant that the constructor argument now also has to be pos_. If you were for some reason generating your own tokens manually (which the wikitables dataset reader was doing) you would need to make the corresponding changes to that code; if you were only creating Tokens using our Tokenizers, then there's no difference to you.

It's quite likely that neither of these changes will affect even a single user, but in theory they could.

List of Commits

baef953 bump version number to v0.8.3
0abefe2 Fix docstrings after inspection (#2655)
a80aac7 Move register to typical location. (#2662)
e1d70bb Add missing paren (#2661)
2bf0779 Fixed ELMO command's lack of encoding specification when reading from… (#2614)
e138d6c TextCat Reader skip_label_indexing Fix (#2653)
6e1ee2e config_allennlp.py uses open(file_path) where file_path is a URL (#2654)
263d340 Upgrade Dockerfile to stretch. (#2647)
fab4b15 Fix quarel explanations (#2648)
37a078a make things backward compatible with spacy 2.0 (#2644)
e79b713 add dependency parser config (#2639)
305bd35 final_state_tuple is a Tuple (#2645)
a4a4306 Checkpointer should check in case user deleted a serialized model (#2531)
3c889ca Update outdated doc (#2641)
12626ac fix sampled softmax tests (#2061)
c06e904 add option "-k" to limit tests in test-install command (#2635)
1357c7e Remove reference to install_requirements.sh from the README (#2633)
0bbd359 Add workaround for missing linking in spacy 2.1, remove install_requirements.sh (#2632)
1b07b48 Bump up spacy version pin to 2.1 (#2626)
e3038a3 bug fixes in drop evaluation and more test cases (#2594)
f8b10a9 Add a no-op trainer. (#2610)
9e72ee0 Fix TextClassificationJsonReader handling of unlabeled instances (#2621)
0106536 Add text classification model (#2591)
43b384d Move some scripts to allennlp/allennlp/tools (#2584)
fe80f9f Fix 'cuda_device' docstring in Trainer.init (#2613)
f19c0ee Enable Pruner class to keep different number of items for different entries in minibatch. (#2511)
3cdb7e2 Ensure contiguous initial state tensors in _EncoderBase(stateful=True) (#2451)
ca998b2 Feature Request: Add a dtype parameter to ArrayField (#2609)
ff90845 change pins to bounds (#2490)
0fffb9b Allow the transition from M to M in the BMES constraint type (#2611)
0542c5a make spacy word splitter return allennlp Tokens (now NamedTuples) by default (#2607)
9e3f405 Only log the keys in the "extras" dictionary when instantiating objects from_params (#2608)
720d306 Handle edge cases in beam search (#2557)
79936e5 Re-use .allennlp when running Docker commands (#2593)
55458f5 fix bugs in naqanet (#2604)
c163b63 Fixed memory error in make_vocab on big dataset. (#2606)
b61d511 context manager that allows predictor to capture model internals (#2581)
3e0fcf0 Update README.md (#2601)
18312a0 Seq2seq dataset reader improvements (#2599)
1adb3e8 Interactive beam search (#2513)
0f7bcf5 Add support for overriding list elements (#2585)
9437b61 disable tutorial test (#2580)
6ea273e Allow checkpointer to be initialized from params (#2491)
b0ea7ab Make tutorial use GPU if available. (#2570)
41174da Fix unit test to work with GPUs. (#2574)
32defc3 fix a bug in augmented_lstm.py (#2534)
cdbac6d Fix min padding length in pretrained NER predictors (#2541)
d0f7170 Make load_archive operate on serialization directories. (#2554)
31af01e Add missing requirement to setup.py (#2564)
c54fcc6 Add NAQANet model for DROP (#2560)
97f3578 add initializer to copynet (#2558)
bbb67e9 Add dataset reader for DROP (#2556)
4d5eade Add official DROP evaluation script (#2559)
3d5560f missing =overrides argument when instantiate Params despite a second time (#2553)
64a8e13 Scope DeprecationWarning errors to just allennlp-internal stuff (#2549)
321cf91 Clarify data_parallel implementation. (#2488)
540ebac Propose a deprecation policy. (#2424)
6d8da97 make archival take an optional output path (#2510)
fefc439 Restore tensorboard epoch metrics to pre-refactoring behavior (#2532)
0205c26 Bump version numbers to v0.8.3-unreleased