KEMT/zpwiki
31
1
Fork 9
Code Issues Pull Requests Releases Wiki Activity

Update 'pages/topics/bert/README.md'

Browse Source
This commit is contained in:
dano 2021-10-21 11:10:46 +00:00
parent 218e472e17
commit 13faa91944
1 changed files with 12 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
pages/topics/bert/README.md
View File

@ -17,17 +17,9 @@ author: Daniel Hládek
- [SlovakBERT](https://github.com/gerulata/slovakbert) od Kinit, a [článok](https://arxiv.org/abs/2109.15254)
- [SK Quad](/topics/question) - Slovak Question Answering Dataset
- bakalárska práca [Ondrej Megela](/students/2018/ondrej_megela)
- https://git.kemt.fei.tuke.sk/dano/bert-train
## Hardvérové požiadavky
[https://medium.com/nvidia-ai/how-to-scale-the-bert-training-with-nvidia-gpus-c1575e8eaf71](zz):
When the mini-batch size n is multiplied by k, we should multiply the starting learning rate η by the square root of k as some theories may suggest. However, with experiments from multiple researchers, linear scaling shows better results, i.e. multiply the starting learning rate by k instead.
| BERT Large | 330M |
| BERT Base | 110M |
Väčšia veľkosť vstupného vektora => menšia veľkosť dávky => menší parameter učenia => pomalšie učenie
## Hotové úlohy
@ -76,3 +68,14 @@ Väčšia veľkosť vstupného vektora => menšia veľkosť dávky => menší pa
- Natrénovať BART model.
- Natrénovať model založený na znakoch.
- Adaptovať SlovakBERT na SQUAD. To znamená dorobiť úlohu SQUAD do fairseq.
## Hardvérové požiadavky
[https://medium.com/nvidia-ai/how-to-scale-the-bert-training-with-nvidia-gpus-c1575e8eaf71](zz):
When the mini-batch size n is multiplied by k, we should multiply the starting learning rate η by the square root of k as some theories may suggest. However, with experiments from multiple researchers, linear scaling shows better results, i.e. multiply the starting learning rate by k instead.
| BERT Large | 330M |
| BERT Base | 110M |
Väčšia veľkosť vstupného vektora => menšia veľkosť dávky => menší parameter učenia => pomalšie učenie