google-research / tapas

End-to-end neural table-text understanding models.
Apache License 2.0
1.13k stars 216 forks source link

Sudden Increase in Loss value while finetuning #173

Open nikithakriz opened 1 year ago

nikithakriz commented 1 year ago

I am finetuning tapas in custom dataset. After creating the dataset while training there is an sudden increase in loss value. After debugging i understood that some of the files creating this issue. But i didn't get the exact logic behind the issue and why that files creating the issue. They are more common like other files. After removing this files from teh dataset the training started with normal loss value.

Issue in Loss Value: Epoch: 0 Loss: 9963.3525390625