Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Sep 12, 2024

• Attempt #3 to add this directory to the new release

  AngledLuffa

  committed
  539760c
  More actions

• Attempt #3 to add this directory to the new release

  AngledLuffa

  committed
  8dd3cb8
  More actions

• Update version -> 1.9.1 to fix missing utils/visualization #1421

  AngledLuffa

  committed
  174768a
  More actions

• Add an init for stanza/units/visualization

  AngledLuffa

  committed
  c96f52e
  More actions

• Set a default for num_layers based on the model type - num_layers=2 seems good for the character classifier

  AngledLuffa

  committed
  b999102
  More actions

• Add a test of the character_classifier

  AngledLuffa

  committed
  7eb3a50
  More actions

• If the seq2seq doesn't predict any spaces in the MWT, use the original word to avoid it going crazy

  AngledLuffa

  committed
  3211e72
  More actions

• Allow --eval_file or --gold_file to override the values used by the run_mwt script to produce the final results

  AngledLuffa

  committed
  15e1408
  More actions

• Allow for loading a save_name for the dev or test sets w/ or w/o the directory

  AngledLuffa

  committed
  c406d43
  More actions

• Save each MWT model based on a flag (useful for debugging individual checkpoints)

  AngledLuffa

  committed
  69d6adf
  More actions

• Refactor build_save_each_filename (just a few lines of code) for later reuse in MWT

  AngledLuffa

  committed
  6d7928a
  More actions

• Add a separate classifier type which checks exactly where MWT splits should go in the original text, rather than using a seq2seq to rebuild the text. For languages such as English where the splits …

  AngledLuffa

  committed
  43466ca
  More actions

• Make it possible for the lemmatizer to return unk characters by reusing the expanded vocab features in the MWT

  AngledLuffa

  committed
  58a1ee0
  More actions

• Refactor the DeltaVocab so as to make it easier for the Lemmatizer to use it

  AngledLuffa

  committed
  ff14d39
  More actions

• Add a test that the delta vocab is working for the MWT

  AngledLuffa

  committed
  6a7fc14
  More actions

• Add the ability to decode unknown tokens from the encoder side of the seq2seq model by temporarily augmenting the vocab with the new tokens. Helps tasks such as MWT which can receive characters out…

  AngledLuffa

  committed
  807987b
  More actions

• Add an option to the seq2seq to never decode unk tokens. This is particularly useful in the MWT where sometimes it was decoding words to look like Bj<UNK>rk<UNK><UNK>ng

  AngledLuffa

  committed
  a8b2754
  More actions

Commits on Sep 6, 2024

• Add a (possibly very noisy) str() for BaseVocab. Actually quite useful in the case of an MWT vocab, for example

  AngledLuffa

  committed
  ddb2caa
  More actions

Commits on Sep 4, 2024

• Separate out a long log line into pieces to make it easier to update going forward

  AngledLuffa

  committed
  8227bd2
  More actions

Commits on Aug 30, 2024

• Remove an unnecessary check - this code path cannot be reached with IN_ORDER_... transition schemes

  AngledLuffa

  committed
  3ff5200
  More actions

• Perhaps 'is' is a bit faster than ==?

  AngledLuffa

  committed
  efd612f
  More actions

• shortcutting len() seems a bit faster, since the TreeStack has its own .length built in

  AngledLuffa

  committed
  532b3ee
  More actions

Commits on Aug 29, 2024

• A few small optimizations to the transition parser

  AngledLuffa

  committed
  038fb67
  More actions

• Default to checking PTB, and add a loop counter so this can be easily used in prof

  AngledLuffa

  committed
  e58ff0e
  More actions

Commits on Aug 25, 2024

• Add a script to run the transition sequences on a constituency dataset. Also useful for speed testing

  AngledLuffa

  committed
  11c4075
  More actions

• Don't start forced oracle transitions until the oracle_initial_epoch

  AngledLuffa

  committed
  1e939d2
  More actions

• Remove code duplication - #1413

  AngledLuffa

  committed
  3de69ca
  More actions

• Add 3.12 as a supported python version now that torch supports 3.12. #1293 (comment)

  AngledLuffa

  committed
  33a5256
  More actions

Commits on Aug 24, 2024

• Add a small bert test including testing an empty word

  AngledLuffa

  committed
  c896431
  More actions

• Sorting by length makes the parser about 10% faster, with most of the speedup coming from better batching of the input encoding. Will need to see if we can better batch the later operations as well

  AngledLuffa

  committed
  a40e84c
  More actions

• Slightly faster - don't use the tokenizer twice unless actually necessary

  AngledLuffa

  committed
  6e48ae0
  More actions

• Refactor the use of the tokenizer - will make it easier to remove unnecessary duplicate uses of the tokenizer when looking for tokens with entirely unknown characters

  AngledLuffa

  committed
  c6be1a1
  More actions

Commits on Aug 18, 2024

• Skeleton for a couple other options for a Hindi model, indic & mbert

  AngledLuffa

  committed
  21d2500
  More actions

• Switch to deeph as the base output name. Ask to download the HI pipeline if needed

  AngledLuffa

  committed
  edd187b
  More actions

Commits on Aug 16, 2024

• Update the tokenizer preparation to include a bunch of adjectives and other words that shouldn't be split. #1410

  AngledLuffa

  committed
  596fcd4
  More actions

Pagination

PreviousNext