Code

Add TinyBERT data augmentation#1923

Code

Merged

MichelBartels merged 21 commits intomasterdeepset-ai/haystack:masterfrom tinybert_data_augmentationdeepset-ai/haystack:tinybert_data_augmentationCopy head branch name to clipboardJan 4, 2022

+259-1

Commits

Commits on Dec 23, 2021

• add tinybert data augmentation

  MichelBartels

  committed
  080b80c

  Copy full SHA for 080b80c

• don't reload glove in tinybert data augmentation

  MichelBartels

  committed
  6543bfc

  Copy full SHA for 6543bfc

• fix unnecessary load_glove call

  MichelBartels

  committed
  40ac53f

  Copy full SHA for 40ac53f

Commits on Jan 3, 2022

• fix type hints

  MichelBartels

  committed
  77f81fd

  Copy full SHA for 77f81fd

• add comments and type hints

  MichelBartels

  committed
  38bfe2d

  Copy full SHA for 38bfe2d

• add batch_size argument

  MichelBartels

  committed
  0aecd64

  Copy full SHA for 0aecd64

• don't predict subwords as alternative for words

  MichelBartels

  committed
  c183702

  Copy full SHA for c183702

• fix subword predictions

  MichelBartels

  committed
  88e2c72

  Copy full SHA for 88e2c72

• limit sequence length

  MichelBartels

  committed
  1b84fd9

  Copy full SHA for 1b84fd9

• actually limit sequence length

  MichelBartels

  committed
  7b6db37

  Copy full SHA for 7b6db37

• improve performance by calculating nearest glove vector on gpu

  MichelBartels

  committed
  62f9565

  Copy full SHA for 62f9565

• add model and tokenizer parameter

  MichelBartels

  committed
  c36ce6d

  Copy full SHA for c36ce6d

• fix type hints

  MichelBartels

  committed
  872600f

  Copy full SHA for 872600f

Commits on Jan 4, 2022

• improve data augmentation performance

  MichelBartels

  committed
  7ec0c6b

  Copy full SHA for 7ec0c6b

• explained limits of script

  MichelBartels

  committed
  fa62cac

  Copy full SHA for fa62cac

• corrected comment

  MichelBartels

  committed
  4ae1cc0

  Copy full SHA for 4ae1cc0

• added data augmentation test

  MichelBartels

  committed
  3c2f08e

  Copy full SHA for 3c2f08e

• don't label every question in augmented dataset as impossible

  MichelBartels

  committed
  8f6f2e7

  Copy full SHA for 8f6f2e7

• add sample glove

  MichelBartels

  committed
  ca64d14

  Copy full SHA for ca64d14

• better handling of downloading of glove

  MichelBartels

  committed
  e9ecdd3

  Copy full SHA for e9ecdd3

• fix typo of last commit

  MichelBartels

  committed
  b33897f

  Copy full SHA for b33897f