Transcribe RTL script to LTR script

[johnlockejrr]

Is there an achievable way to transcribe a RTL script (Avestan) to a LTR script (Latin)? If so, how can be done? I don't see an option out of the box.

[mittagessen]

On 25/01/01 04:10AM, johnlockejrr wrote: Is there an achievable way to transcribe a RTL script (Avestan) to a LTR script (Latin)? If so, how can be done? I don't see an option out of the box.

You could teach a standard recognition model to output the LTR transliteration for the input but that doesn't work particularly well in my experience as the structure of the CTC loss requires the network to 'store' the recognition results internally to then emit them reversed into the output. The new recognizer doesn't have this issue as it isn't trained with CTC. There will be a new base model for that one by next week.

[johnlockejrr]

Perfect! Looking forward!

Note: Will you add the model to kraken or to party?

[johnlockejrr]

Any update on this?

[mittagessen]

On 25/01/09 01:51PM, johnlockejrr wrote: Any update on this?

The model is up now. Instructions are in the [party](https://github.com/mittagessen/party) repository.

[johnlockejrr]

For segmentation, is the new blla model on zenodo? Or the Pretrained multilingual Party model (that seems to be a recognition model)?

[johnlockejrr]

Can I try to finetune the model with a NVIDIA GeForce RTX 4070 12Gb? Worked fine for kraken models.

[mittagessen]

The new segmentation model isn't there yet because it didn't converge for whatever reason so I need to debug that. The pretrained party model is the right one for the new recognizer but it will also download it for you if you run the inference stuff without an explicit model given. Right now you'll need ~20Gb of memory for fine-tuning but I'll add some small changes later today that will reduce memory consumption. You can run inference with large batch sizes on that GPU though.

…

On 25/01/09 10:51PM, johnlockejrr wrote: Can I try to finetune the model with a `NVIDIA GeForce RTX 4070 12Gb`? Worked fine for kraken models. -- Reply to this email directly or view it on GitHub: #669 (comment) You are receiving this because you commented. Message ID: ***@***.***>