FineTuning Segmentation detecting different structure than active

[nostalgic-math-student]

Hello, I have found an issue while trying segtrain finetuning.

My data is ALTO XML files and image to back the text, I would like to do segmentation tasks with segtrain models.
So I have this batch of models which I would like to finetune with an extra batch of data in the same ALTO Format SCHEMA version 4.2
However when I try to run ketos segtrain I get this error:

ValueError: Training data and model class mapping differ (bl: set(), regions: {'composed'}

Where I have checked every file has the same structure and format (I ran the parser by my own) but Kraken check warns of "regions composed" not present in the files.

Weird enough matter, if I try to do "fineTune" using the same training data I don't get this error again, meaning there is something I may have missed.

Thanks for the help, hope this is not an error but might be something.

[mittagessen]

On 25/11/20 01:30PM, nostalgic-math-student wrote: nostalgic-math-student created an issue (mittagessen/kraken#746) Hello, I have found an issue while trying segtrain finetuning. My data is ALTO XML files and image to back the text, I would like to do segmentation tasks with segtrain models. So I have this batch of models which I would like to finetune with an extra batch of data in the same ALTO Format SCHEMA version 4.2 However when I try to run ketos segtrain I get this error: ``` ValueError: Training data and model class mapping differ (bl: set(), regions: {'composed'} ``` Where I have checked every file has the same structure and format (I ran the parser by my own) but Kraken check warns of "regions composed" not present in the files. Weird enough matter, if I try to do "fineTune" using the same training data I don't get this error again, meaning there is something I may have missed.

Can you show me which commands you're running exactly? I guess you're not defining explicit training/validation/test splits so you might get different files in them between different runs which would explain the intermittancy of the error. That's a hint to use fixed splits ;). The region label comes from a fallback mechanism that derives its label from the XML element name, in this case ComposedBlock, if a region defines neither a TYPE nor TAGREFS. Look for one of those and if you don't want to train on it just filter it out with the `--valid-regions` option.