Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Apr 20, 2024

• same thing, converting VIT with Stanza 1.8.x should already be using a recent CoreNLP

  AngledLuffa

  committed
  6e442a6
  More actions

• assuming >4.5.1 is the standard for our unit tests at this point

  AngledLuffa

  committed
  f380dd7
  More actions

• Bump Stanza version number for new release

  AngledLuffa

  committed
  0f7f2af
  More actions

• All models now save/resave the finetuned transformers

  AngledLuffa

  committed
  98455fc
  More actions

• Add peft wrapping & saving to POS, complete with potentially using unique names with multiple peft

  AngledLuffa

  committed
  e829f4f
  More actions

• Separate out evaluate_trainer in the depparse to make it easier to test an already-loaded model

  AngledLuffa

  committed
  a3de81e
  More actions

• Reuse the peft code from NER in the depparse

  AngledLuffa

  committed
  b7e43d1
  More actions

Commits on Apr 18, 2024

• use_reentrant in gradient_checkpointing avoids some mysterious negative interaction between peft and gradient_checkpointing. thanks for the tip @Jemoka

  AngledLuffa

  committed
  597d48f
  More actions

Commits on Apr 16, 2024

• Bugfix - with retag_pipeline None, the evaluation doesn't work w/o a default value for retagged_treebank

  AngledLuffa

  committed
  122f16e
  More actions

Commits on Apr 14, 2024

• Separate out an evaluate function which takes a loaded model, making it possible to load those models elsewhere and then get an NER score

  AngledLuffa

  committed
  35a5a9e
  More actions

• Add an option to make a maximum batch size in terms of #words for NER. Helps avoid OOM when finetuning bert or using peft

  AngledLuffa

  committed
  b11d152
  More actions

• Use a specialized peft name for the NER peft. Will make it easier to integrate with the Pipeline

  AngledLuffa

  committed
  fb2ccb1
  More actions

• Rough draft of multiple peft names for different processors

  AngledLuffa

  committed
  71212f6
  More actions

• Reuse a repeated block of code - can maybe do the same for other modules which use transformers

  AngledLuffa

  committed
  d402043
  More actions

• Always make new copies of the transformer when training, including when using peft

  AngledLuffa

  committed
  d172f66
  More actions

• Use a named peft adapter on the constituency parser, similar to the classifier. Will need to add the capability of loading multiple peft adapters on the same transformer

  AngledLuffa

  committed
  4304fc5
  More actions

• Save & load the peft adapter in the classifier using a specific adapter_name. Will hopefully make it easier to switch adapters once we extend that to other annotators

  AngledLuffa

  committed
  0199297
  More actions

Commits on Apr 11, 2024

• When --log_norms is set, log a bit less frequently - otherwise, just way too noisy. Also, log when evaluating

  AngledLuffa

  committed
  9ae98b1
  More actions

• Fix typo

  AngledLuffa

  committed
  52e8990
  More actions

Commits on Apr 9, 2024

• Fix a typo

  AngledLuffa

  committed
  fc583c6
  More actions

• Name change the test_ensemble, since it doesn't test other features we will need to test of the ensemble

  AngledLuffa

  committed
  8f318a1
  More actions

• Rename and move the function to drop singleton trees as well

  AngledLuffa

  committed
  fbb1d12
  More actions

• Refactor a couple more functions into constituency/utils

  AngledLuffa

  committed
  b864729
  More actions

• Minor change - fewer direct uses of model

  AngledLuffa

  committed
  77f0f53
  More actions

• Move bulk_apply into the model - will be easier to refactor the training loop or other parts

  AngledLuffa

  committed
  c783526
  More actions

• Match the interface for bulk_apply with a fail=... parameter

  AngledLuffa

  committed
  55ce6a8
  More actions

• Refactor some utility methods and make a MultiState which holds multiple states at once for the Ensemble. Should make it easier to refactor parse_sentences() and ultimately integrate the Ensemble w…

  AngledLuffa

  committed
  4f88ac3
  More actions

• Refactor State into a separate module

  AngledLuffa

  committed
  ba2d27c
  More actions

Commits on Apr 8, 2024

• Add a test of constituency parser's ensemble.py which checks that the main inference loop works when using multiple of the same parser, at least

  AngledLuffa

  committed
  ce6a246
  More actions

• turn on this test on travis

  AngledLuffa

  committed
  efe7b6b
  More actions

• Condense the usage of initial_state_from_preterminals - can include the gold sequence as well

  AngledLuffa

  committed
  afd1b62
  More actions

• Slightly rearrange in the training loop - mission is to make it easier to refactor everything

  AngledLuffa

  committed
  a1f3483
  More actions

• Add an option to train the depparse w/o upos... can test just a transformer, for example

  AngledLuffa

  committed
  15b136b
  More actions

• Sanity check the gold doc before running the scoring. Otherwise, the scorer produces a very strange error for a None dependency

  AngledLuffa

  committed
  909159c
  More actions

• Make reverse_sentence a property rather than a method

  AngledLuffa

  committed
  9efe95c
  More actions

Pagination

PreviousNext