Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Mar 12, 2024

• Update get_tqdm.py

  SajanShetty

  authored
  b296934

Commits on Mar 1, 2024

• Will update Stanza version to quickly fix a few bugs

  AngledLuffa

  committed
  c2d72bd

• Use a get() to avoid crashing if an older model with no bert_funetune set is loaded. Addresses stanfordnlp#1357

  AngledLuffa

  committed
  13ee3d5

• Add some debug logging when building a retag_pipeline - goal is to make sure it's calling the correct POS model

  AngledLuffa

  committed
  6e2520f

• Allow for an individual pipeline to override which device it is placed on. Furthermore, this avoids a TypeError in the event someone tries to do that. First spotted in stanfordnlp#1089

  AngledLuffa

  committed
  44058a0

Commits on Feb 25, 2024

• Quieter logging when building a peft wrapper

  AngledLuffa

  committed
  17eb6fc

Commits on Feb 24, 2024

• Fix TOP_DOWN parser for da_arboretum, which needs to look at the actual root labels rather than just assume ROOT

  AngledLuffa

  committed
  e89a7d4

• Add the ability to process a few languages with prepare_resources.py

  AngledLuffa

  committed
  5f18a61

• Update sentmient to also have charlm & transformer versions. The transformer versions actually use peft... should update run_sentiment to have an option for that

  AngledLuffa

  committed
  363bbec

• Minor logging / typo improvements to conparser

  AngledLuffa

  committed
  4a7052b

• Initial attempt to chop up long inputs to a transformer into pieces that the transformer can digest, even if it isn't necessarily going to give great results for the later tokens in the sentence. A…

  Show description for 3e21404
  AngledLuffa

  committed
  3e21404

• Add a script which filters silver trees by the variance of their scores over a sequence of trained models. Works either by taking the least or the most variance. Hopefully this script will work to …

  Show description for fe45f11
  AngledLuffa

  committed
  fe45f11

• Only do args['num_generate'] once... possibly fixing a bug where num_generate was already expanded

  AngledLuffa

  committed
  187850c

• Update constituency evaluation to accommodate the per-tree f1

  Show description for 8a6c543
  AngledLuffa

  committed
  8a6c543

• Add an option to turn off saving the optimizer when saving each constituency model

  AngledLuffa

  committed
  837dd03

• Add flags to control how often the save_each models get saved. Intent is to make it so that we can skip 10 models at a time and use a series of these to measure the variance in a silver dataset's s…

  Show description for 56372f5
  AngledLuffa

  committed
  56372f5

• Remove unnecessary repetition of existing save_each argument

  AngledLuffa

  committed
  5cdc8ed

• Refactor a couple parsing functions in trainer.py Use this to build a new silver dataset producing script which uses two ensembles at once to directly find the matching trees along with counting th…

  Show description for 9edcad9
  AngledLuffa

  committed
  9edcad9

Commits on Feb 23, 2024

• Bump version number, including models version number, as the new models will include peft for at least constituency and sentiment. NER and depparse are also available, but have not show any improve…

  Show description for d4aeba9
  AngledLuffa

  committed
  d4aeba9

Commits on Feb 22, 2024

• Whitespace

  AngledLuffa

  committed
  581837f

• Set the trainer logging level to DEBUG when starting a constituency training run. Switch out a bunch of training log calls to use the training logger specifically to make it easier to log things li…

  Show description for 64d7f68
  AngledLuffa

  committed
  64d7f68

• Add a block of debug logging to get_optimizer ... the constituency parser used to report what optimizer it was creating, and I miss those log lines

  AngledLuffa

  committed
  6ac469d

• Integrate peft with the constituency parser.

  Show description for f645a8c
  AngledLuffa

  committed
  f645a8c

• Add bert_finetune_layers to depparse as well (still OOMs on normal batch sizes)

  AngledLuffa

  committed
  766b0db

• Refactor building the optimizer in conparse to use the utils get_optimizer. Should make it easier to integrate a peft wrapper to the conparser

  AngledLuffa

  committed
  9eb0c74

Commits on Feb 21, 2024

• Add a test that peft finetuning, saving, and loading works. Include a note on something we need to do in the bert test

  AngledLuffa

  committed
  6b5b019

• Hide imports inside the methods which use them, so as to make peft not a requirement yet

  AngledLuffa

  committed
  ab13e03

• peft and two stage optimizer

  

  Jemoka

  authored and

  AngledLuffa

  committed
  0126a41

• NER weights and biases gradient tracking

  

  Jemoka

  authored and

  AngledLuffa

  committed
  ff219ea

• Add AdaBelief and mirror_madgrad as alternate optimizer options

  AngledLuffa

  committed
  8a82fce

• Don't add bert parameters at all in the single optimizer peft case if the bert_learning_rate is 0

  AngledLuffa

  committed
  131b41e

Commits on Feb 18, 2024

• Refactor load_doc so that the determinstic portion of the trainer can use text with goeswith eliminated as well

  AngledLuffa

  committed
  191a05f

• Report how many sentences get loaded from the handparsed treebank

  AngledLuffa

  committed
  8908e95

• Remove words that are part of a goeswith phrase when training a lemmatizer, since some treebanks have the standard of making the lemma the complete goeswith phrase, and that works pretty horribly f…

  Show description for 9d44885
  AngledLuffa

  committed
  9d44885

• Turn methods which really are just static methods into @staticmethod

  AngledLuffa

  committed
  8801c55

Pagination

PreviousNext