Piotr Czapla
8733487d55
Make the validate vs train decision based on the existance of cls_last.pth istead of a model directory
2019-02-15 01:17:53 +01:00
Piotr Czapla
5dced1e488
Remove bidir
2019-02-15 01:16:37 +01:00
Piotr Czapla
0e6534ad7b
Expose num_lm_epochs in ulmfit eval
2019-02-15 01:11:39 +01:00
Piotr Czapla
0f084168c1
Merge branch 'master' of https://github.com/n-waves/ulmfit-multilingual
2019-02-14 22:35:29 +01:00
Piotr Czapla
cd47b3b5dc
Fix use_moses=True for mldoc so that it is identical to wiki with uses_moses=False
...
The issue was that Moses was executed after pre_rules when use_moses = True, But when data set was pre tokenized with Moses (use_moses=False) the pre_rules were executed after.
So our wikipedia had the following processing:
- raw text
- Moses
- pre_rules
- split(' ') # fastai BaseTokenizer
- post_rules
- sentence piece
While mldoc had the following tokenziation
- raw text
- pre_rules
- Moses
- post_rules
- sentence piece
After fix I've retrained the classfiers (without finetuning) and I haven't notice huge changes in the performance. 4 languages received slight improvment 4 got a slight decrease in performance.
2019-02-14 22:35:20 +01:00
Piotr Czapla
c28c0fde16
Make ulmfit eval more secure and give more flexibility in dataset_template
...
The dataset_template can use lang as additional token to construct globs patterns.
2019-02-14 22:28:25 +01:00
Marcin
fdac9f7ccd
Save only the best LM model
2019-02-14 14:01:33 +01:00
Marcin
b609951561
Fix path of pretrained model
2019-02-14 00:00:56 +01:00
Piotr Czapla
1340f4235c
Improve ulmfit eval to allow for zeroshot laser evaluation
2019-02-13 15:29:55 +01:00
Marcin
0e2211dfaf
Merge branch 'master' into use-configs
2019-02-13 14:10:59 +01:00
Marcin
2eee051c67
Expose max length parameter
2019-02-13 13:40:05 +01:00
Piotr Czapla
3586a7dcf9
Merge branch 'pr/29'
2019-02-12 20:43:44 +01:00
Piotr Czapla
e7271f2a29
Add ability to evalulate multiple models at once
2019-02-12 15:01:01 +01:00
Marcin
8497cc1e5c
Move LM and classifier parameters to configs
2019-02-12 03:09:51 +01:00
Piotr Czapla
a630242f97
Expose validate_cls in ulmfit module
2019-02-11 11:01:08 +01:00
Piotr Czapla
c63fe258d2
Fix validataion and add option to add noise to training labels
2019-02-11 10:53:25 +01:00
Piotr Czapla
26736d95de
Add code to test & train mldoc classifier
2019-02-10 09:52:58 +01:00
Tomasz Pietruszka
6fa30829c4
Merge branch 'master' into backwards-lm
2019-01-14 17:20:48 +01:00
Tomasz Pietruszka
5ba8ae4c59
non-ascii char removed from code. Caused display bugs
2019-01-13 17:29:55 +01:00
Tomasz Pietruszka
b00410e0cb
LM save with_opt fix
2019-01-13 17:28:39 +01:00
Tomasz Pietruszka
a1e66c39d4
tokenzier->tokenizer typo
2019-01-13 17:27:29 +01:00
Tomasz Pietruszka
593c5661bf
Added alpha and beta params for RNNTrainer
2019-01-13 17:26:16 +01:00
Tomasz Pietruszka
e6407fe0c9
Added the param and model type for BwdLM
2019-01-10 00:47:22 +01:00
Piotr Czapla
0085c18ae0
Fix splitting by article for local languages and make it more memory efficient
...
The issue was that the code assumed that empty lines have space followed by a new line, which isn't the case for datasets generated by our scripts.
2019-01-03 12:16:41 +01:00
Piotr Czapla
f784d7bcd2
Make the article detection code work with our wikitext
2019-01-01 15:13:14 +01:00
Piotr Czapla
0945c699c7
Fixes #25 by adding article title in markdown format to wiki text
2019-01-01 15:07:02 +01:00
Piotr Czapla
4a858f3572
Simplfy and unify input data parsing
2019-01-01 14:43:12 +01:00
Piotr Czapla
82d6a30b11
Fix SentencePiece implementation, to get 94.5% on imdb
...
- train on whole articles when tokenizer is sentencepice
- add moses tokenizer to get the same tokens on imdb as on wt103
- apply pre / post processing rules to moses tokenzier so that sentencepiece works on correctly preprocessed text (lowercased with html removed)
- add alpha implementation on xnli (no tests)
- remove vocab adaptation when swiching from wiki to imdb. As otherwise we get 50% of missing words and 93% accuracy on imdb
2018-12-27 15:14:41 +01:00
Piotr Czapla
ff30fb5642
Fastai upgrade
2018-12-27 14:57:15 +01:00
Piotr Czapla
f9394b9af1
Add callbacks to save history and best weights remove bs & drop_mult
2018-12-12 00:31:16 +01:00
Piotr Czapla
1b9d04d7ef
Expose use_test_for_validation as param to train
2018-12-09 22:45:21 +01:00
Piotr Czapla
34b1d18600
load wikipedia as articles for v & fv tok.
2018-12-09 22:42:17 +01:00
Piotr Czapla
2338618563
Make biclasifier head a hyperparameter.
2018-12-09 00:39:27 +01:00
Piotr Czapla
978bbcc923
Fix issue with finetuning language model (it wasn't freezed)
2018-12-08 23:10:31 +01:00
Piotr Czapla
50ab34eea9
Fix classification scripts and give a way to test accuracy on test set
2018-12-07 16:10:20 +01:00
Piotr Czapla
4ae2a158a2
Respect batch size in training lm model.
2018-12-06 23:29:43 +01:00
Piotr Czapla
2acbf15555
Tweak hyper training params of cls (drop_mul, bs, true_wd=True)
...
I've set the same hyperparams as in lesson3
2018-12-05 16:27:14 +01:00
Piotr Czapla
4454af167f
Use more text during pretraining of imdb
2018-12-05 16:26:16 +01:00
Piotr Czapla
5524006d81
Respect max_vocab
2018-12-04 16:41:22 +01:00
Piotr Czapla
908c3d7e8a
bug fix
2018-12-04 16:41:07 +01:00
Piotr Czapla
f25deb3049
Use trn + tst for LM training
2018-12-04 01:43:01 +01:00
Piotr Czapla
9826f6881c
Fix the way trn & val set is created in imdb
2018-12-04 01:35:59 +01:00
Piotr Czapla
039624870a
Fix loading tokenized data set in train cls
2018-12-04 01:28:03 +01:00
Piotr Czapla
a499bf9a20
Make cls train work with relative paths
2018-12-04 01:24:44 +01:00
Piotr Czapla
35a5dcb75a
Fix issue when running cls training from command line
2018-12-04 00:59:36 +01:00
Piotr Czapla
82c955ce6a
Add different tokenization algorithms to train_clas
2018-12-04 00:51:21 +01:00
Piotr Czapla
83427aadc6
Add moses with fastai preprocessing
2018-12-02 21:43:57 +01:00
Piotr Czapla
b9eb7388f6
Fix bidir for fastai tokenizer
2018-12-01 23:58:14 +01:00
Piotr Czapla
0c4aed6d05
Really fix conversion from str to Tokenzier
2018-12-01 16:51:21 +01:00
Piotr Czapla
e14c967cc8
Fix string parsing in tokenzier
2018-12-01 16:46:25 +01:00
Piotr Czapla