On wikitext-103 the model trains in Ā±18h on 1080TI
100k is huge, it makes it hard for model to learn useful relations between words for Russian you may want to use SentencePiece with 25k tokens, it works really well for Polish (better than sentence piece with 50k tokens, way better than 100k tokens).
You may check our paper & presentation there is an example that show how a different number of tokens influence the way a random sentence is being split.
looks like the english wikipedia dump will be 25-27 mio sentences when i have finished the script to remove āabnormal sentenceā. From my measurements one epoch will take 20 hours.
Iāve also trained a language model and classifier for Hindi, achieving a perplexity of ~35 on 20% validation set of 55k Hindi Wikipedia articles. Iām using Fastai v1 and Sentencepiece for Tokenization. I would like to compare our models on the BBC News classification dataset. Would you mind sharing your score?
@disisbig can you make a thread for you language and put it into the top entry? Re comparison we are in process of assembling the language models in one repository to ensure reproductability. https://github.com/n-waves/ulmfit-multilingual Do you want to contribute your lm and hyper Parmas?
Thanks @piotr.czapla. Iāve created the threads for Hindi and Punjabi. Iāll soon raise a PR to contribute my models and hyper-params to ulmfit-multilingual
Folks, would anyone know if one can use a language model (instead of word vecs) for sequence 2 sequence translation? Think Jeremy mentioned that in previous deep learnng part II in lesson 11 where he demoed translation wird word vecs.
Not sure I got this correct and its possible, pointers welcome.
the breakthrough in this paper is that it is not a RNN. RNNs takes a long time to train and have issues with translating long sentences. I have been training RNNs where it tok 15 hour to process 10 epochs on 2.5e8 tokens. The awd_lstm rrn in fastai is very interesting as a model it just requires a lot of patience to train
The same perplexity/accuracy can be reached in about an hour using the transformerXL @sgugger implemented recently. it handles long sentences much more elegantly (attention mechanism) and can be parallelized.
In short - if you want to train languagemodels for translation or classification etc. you will do it faster an better using the transformerXL model.
If you manage to do that, please tell me how. Those models are heavy and require a much longer time to train! Those itās true it takes less epochs to reach a ppl as low as the AWD-LSTM on WT103, it still takes more compute time.
thx @Kaspar got it ! I will take a look at it then for a pet project I want to do.
I am still looking for simple example code (afaict there are no examples in fast.ai notebooks) how to use a language model for translations, I just saw fast.ai examples from last term part II (lesson 11) using word vecs. @sgugger