Neural network has been trained on mode 1 images (not really an issue)

[johnlockejrr]

I'm fine-tuning a model with kraken and I get this warning at the beginning:
[04/24/24 17:56:30] WARNING Neural network has been trained on mode 1 images, training set contains mode L data. Consider setting 'force_binarization'

Should I convert my groundtruth images to mode 1, force_binarization or just ignore?

[mittagessen]

On 24/04/24 10:58AM, johnlockejrr wrote: I'm fine-tuning a model with kraken and I get this warning at the beginning: `[04/24/24 17:56:30] WARNING Neural network has been trained on mode 1 images, training set contains mode L data. Consider setting `force_binarization`` Should I convert my groundtruth images to `mode 1`, `force_binarization` or just ignore?

You should probably just ignore it, especially if the model has been trained with a kraken version <5 which had a bug that caused the 1 byte per channel mode not to be correctly promoted with multiple dataloader threads. If the model *has* indeed been trained on binarized only data the fine-tuned one will not work well on binary data so depending on the use-case you might consider binarization your training data. But there isn't much reason nowadays for binarized-only models and it introduces an additional processing step (and therefore error source) in the pipeline.

[johnlockejrr]

Perfect! Thank you for clarification, I tested the model, works flawlessly.

…

On Wed, 24 Apr 2024 at 23:05, mittagessen ***@***.***> wrote: On 24/04/24 10:58AM, johnlockejrr wrote: > I'm fine-tuning a model with kraken and I get this warning at the beginning: > `[04/24/24 17:56:30] WARNING Neural network has been trained on mode 1 images, training set contains mode L data. Consider setting `force_binarization`` > > Should I convert my groundtruth images to `mode 1`, `force_binarization` or just ignore? You should probably just ignore it, especially if the model has been trained with a kraken version <5 which had a bug that caused the 1 byte per channel mode not to be correctly promoted with multiple dataloader threads. If the model *has* indeed been trained on binarized only data the fine-tuned one will not work well on binary data so depending on the use-case you might consider binarization your training data. But there isn't much reason nowadays for binarized-only models and it introduces an additional processing step (and therefore error source) in the pipeline. — Reply to this email directly, view it on GitHub <#595 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD44GHU7QZXM7AHAQGJ2JL3Y7ANDHAVCNFSM6AAAAABGXMLL5CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDANZVHA2DANRYGM> . You are receiving this because you authored the thread.Message ID: ***@***.***>