Layoutlmv2 - Document Classification predicting same class always about transformers-tutorials HOT 9 CLOSED

Always use 5e-5 to fine-tune Transformer-based models ;)

from transformers-tutorials.

thanks for your reply.
I had a balanced dataset . I have solved the issue. The problem was the learning rate . I changed different learning rates and trained for longer epochs and now the model seems to predict different classes.

from transformers-tutorials.

You can leverage LayoutLMv2FeatureExtractor for LayoutXLM (as the image-related part is the same), however, the tokenizer for LayoutXLM is based on XLM-RoBERTa. So you would have to use the following tokenizer:

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutxlm-base")

So you would have to use both the feature extractor and tokenizer separately, there's no processor defined for it, for now.

from transformers-tutorials.

The model's training accuracy was above 90% after certain epochs , but it is always predicting the same class.

So you do have a very imbalanced dataset (i.e. a lot more samples of a given class compared to another class)?

from transformers-tutorials.

the issue was solved and closed

from transformers-tutorials.

yes sure ...thanks...

I have another question.

Can you please tell me how to load Layoutlxlm model (multilingual) ?

What is the tokenizer , feature extractor and processor for Layoutxlm ?

from transformers-tutorials.

Thank you for the clear explanation @NielsRogge

from transformers-tutorials.

Hai @Jerome-Michael
I am also facing the same issue can you please share what learning rate you used and how many classes you have in your dataset and how you balanced the dataset.

Thanking you in advance.

from transformers-tutorials.

Hi @Laxmi530 .
I think I used 1e-5 . I trained for little more epochs around 50. I had 3 classes. totally balanced

from transformers-tutorials.