0

I'm working now on multi-label classification service and currently I have prepared training dataset. The problem is that current training dataset has a lot of different labels (around 80 000).

At first I tried BERT from Google through Simpletransformers: https://github.com/ThilinaRajapakse/simpletransformers#minimal-start-for-multilabel-classification

But the problem is that you need there for each input binary matrix with labels so that means I need for each keyword to train huge binary matrix with length 80 000. So it's not usable.

Then I tried Ludwig: https://uber.github.io/ludwig/examples/#multi-label-classification

But after first test it seems like Ludwig remember only first 10 000 most frequent labels so it's not usable for my case too.

Now I'm trying it with FastText and it seems it works well, just need a lot of CPUs and computational time.

My question is if you think FastText is good option for my case or if you can recommend me something else when I do multi-label classification with so much labels.

Thanks!

fliprs
  • 47
  • 1
  • 6

1 Answers1

0

fasttext is a good option. You can use Google collab. In spite of more CPUs, you can select runtime -> change runtime type -> GPU/TPU.

To use fasttext in Google collab, you can start with following code snippet:

!wget https://github.com/facebookresearch/fastText/archive/{version}.zip
!unzip {version}.zip
%cd fastText-{version}
!make
MrRaghav
  • 335
  • 3
  • 11