wassname/multifit
1
0 Fork 0
mirror of https://github.com/wassname/multifit.git synced 2026-10-29 13:20:17 +08:00
Code Issues Packages Projects Releases Wiki Activity
284 Commits 17 Branches 0 Tags
59e9966fdcfc953ef322c341516c8763c9c9b140
Commit Graph
284 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Piotr Czapla 59e9966fdc Update to README.md with resutls and links to example notebooks 2019-11-06 11:32:11 +01:00
Piotr Czapla 400e54dec1 Compatiblity with fastai 1.0.47 2019-11-06 10:45:48 +01:00
Piotr Czapla d85bd94241 README + requirements 2019-11-06 10:45:18 +01:00
Piotr Czapla 30993f81a3 Fix command line training & rename _learner to get_learner 
The learner has good use as a public method so there is no point in keeping it protected
 2019-11-06 10:44:29 +01:00
Piotr Czapla 8f46753195 Fix the weight decay - issue that was causing lower results on DE. 2019-11-02 09:35:13 +01:00
Piotr Czapla a47b20fcf5 Add option to turn off label smoothing eps normalisation & weight convertion New weights -> Old format 2019-11-02 09:34:40 +01:00
Piotr Czapla a4b06149b1 Expose method to get configured text processor 2019-11-02 09:32:53 +01:00
Piotr Czapla e009a02aab Update the paper_version configuration to match more closely original experiements 
- the dropouts values are changed in new fastai, we restore the orginal values
- batch size 18 in classifier works a bit better with the learning rate
- add support for gradient clipping and set it to 0.12
- divide label smoothing eps by number of classes
Further differences that werent fixed:
- different data loader and sampler the batches are a bit different (longer?)
- use of masking on padding tokenss
 2019-10-21 10:18:52 +02:00
Piotr Czapla 3c2364e3d8 Clean up bugs caused by the new tokenizer class 2019-10-21 10:04:59 +02:00
Piotr Czapla 031e0c18eb Add diff printing during load_ 2019-10-15 16:55:15 +02:00
Piotr Czapla 70c74a1cc5 Refactor tokenization 
So that we can instantiate tokenization before we know what dataset we want to use it on. Previously it was tidly copuled.
 2019-10-15 04:17:54 +02:00
Piotr Czapla 77a2780a6b Clean up datasets 2019-10-15 00:57:40 +02:00
Piotr Czapla ad50c73377 Make ULMFiT name consistant. 2019-10-15 00:12:31 +02:00
Piotr Czapla 43f5c1282d Add sotabench scripts 2019-10-15 00:09:18 +02:00
Piotr Czapla f846bf8a4b Add OpenFIle preproc to SentencePiece preproc 2019-10-14 17:20:26 +02:00
Piotr Czapla b362f2a7cb from_pretrained 2019-10-14 17:18:28 +02:00
Piotr Czapla 5a3fbcecb5 Ability to use fastai databunch directly 2019-10-14 17:18:13 +02:00
Piotr Czapla 6336f5ebb8 Make it possible to create databunch out of dataframes 2019-10-14 13:14:07 +02:00
Piotr Czapla 91dbd9bc84 Clean up configurations 
So we use the name of function automatically
 2019-10-14 13:13:09 +02:00
Piotr Czapla 7847331751 Remove cupy from requirements.txt 
qrnn isn't using it anymore
 2019-10-14 13:11:52 +02:00
Piotr Czapla 1fe2bd46a6 finall clean ups 2019-10-07 14:50:03 +02:00
Piotr Czapla 3fe6c19af2 fix spelling eeror in finetune_lm 2019-09-08 08:43:15 +02:00
Piotr Czapla cac4df27bd Add paper version configuration and orignal ulmfit 2019-09-07 22:37:26 +02:00
Piotr Czapla c390b4ed35 Add code to JA-multifit 2019-09-07 21:47:43 +02:00
Piotr Czapla dd473aa98d Add repoduction of Multifit result for JA using newest hyper params 2019-09-07 21:18:07 +02:00
Piotr Czapla 1ff9e01766 Remove unused code 2019-09-07 21:17:32 +02:00
Piotr Czapla ee7c23a3be Remove old multifit logs 2019-09-07 19:12:44 +02:00
Piotr Czapla 02ee52d0ef Update README.md 2019-09-07 19:11:52 +02:00
Piotr Czapla 26e54a9c7d Use the pervious datset_path from finetuning for classsificator training 2019-09-07 19:11:29 +02:00
Piotr Czapla 2cdf380adf New bolerplate code 2019-09-07 16:19:09 +02:00
Piotr Czapla 8162cc5648 Clean up the old multfit training code 2019-09-07 15:22:01 +02:00
Piotr Czapla 2fe5c8a588 Update to fastai v1.0.57 - use new sentence piece implementaiton & sizes of hidden layers 2019-08-29 15:51:17 +02:00
Piotr Czapla 68d6b1c829 fix bug in the hack for poleval reddit 2019-06-10 20:31:04 +02:00
Piotr Czapla a7ac4f5170 missing file 2019-06-10 20:14:57 +02:00
Piotr Czapla 7d59c4ecf2 Sentence piece has new option -fix that changes coverage to 99.95% 2019-06-10 20:14:11 +02:00
Piotr Czapla ce0a29f385 Make it possible to set ftseed in poleval19_init 2019-06-06 17:45:13 +02:00
Piotr Czapla 854391a130 Merge branch 'reproduce-poleval' of https://github.com/n-waves/ulmfit-multilingual into reproduce-poleval 2019-05-19 00:17:27 +02:00
Piotr Czapla 38e131f802 Make the ulmfit compatible with old fastai (ulmfit_multilingual) 2019-05-19 00:17:03 +02:00
Marcin 7326df2c98 Train sentencepiece on unsup 2019-05-19 00:12:34 +02:00
Piotr Czapla 6f4db9de1a Fix ensemble output (by removing index column) 2019-05-18 22:50:55 +02:00
Piotr Czapla 06c32ae2dc Add ability to create folders when saving ensemble output 2019-05-18 21:59:12 +02:00
Piotr Czapla f38f8c0670 fix issue in the databunch name generation 2019-05-18 21:48:32 +02:00
Piotr Czapla 8927e45fb2 Add ensemble commnad 2019-05-18 21:48:08 +02:00
Marcin 15c2d0663a Save predictions for a proper dataset 2019-05-18 14:42:35 +02:00
Piotr Czapla 7a33ea5d1f Add no-test + ability to start training from the classifcation data set without wiki 2019-05-18 13:10:29 +02:00
Piotr Czapla 75b934bb83 Merge branch 'reproduce-poleval' of https://github.com/n-waves/ulmfit-multilingual into reproduce-poleval 2019-05-15 18:00:45 +02:00
Piotr Czapla 7c9129670e Poleval 19 ablation experiments 2019-05-15 18:00:40 +02:00
Marcin 30f5a7901c Save predictions 2019-05-14 15:34:40 +02:00
Piotr Czapla 7955db37ad Merge branch 'reproduce-poleval' of https://github.com/n-waves/ulmfit-multilingual into reproduce-poleval 2019-05-14 13:27:03 +02:00
Piotr Czapla ad3c11edb8 Add ability to turnoff weighted CrossEntropy 2019-05-14 13:26:57 +02:00