Request For Adding ParSeq - text recognition model

[nikokks]

🚀 The feature

Hello,

I mainly use the text detection and text recognition models with your framework.

As I have seen: the most recent models that you propose in text recognition, namely MASTER and SAR, are not yet operational.

However at the text recognition level, there is a recent model that gets very impressive performances: PasrSeq.

Here are the references:
https://github.com/open-mmlab/mmocr/blob/main/configs/textrecog/master/README.md
https://github.com/open-mmlab/mmocr/blob/main/configs/textrecog/abinet/README.md
https://paperswithcode.com/paper/scene-text-recognition-with-permuted#code
https://github.com/baudm/parseq

Would it be possible to add this recognition text to the models you propose?

Thanks a lot for your work !

Motivation, pitch

I'm working with text recognition models and a recent model in state of the art outperforms on all test datasets.
I would like use this model with your framework pipelines

Alternatives

No response

Additional context

No response

[felixdittrich92]

Hi @nikokks 👋 ,
fyi MASTER and SAR are fixed in both frameworks on main branch will be released soon with 0.5.2

I agree with the ParSeq addition firstly i have had in mind to add ViTSTR but yes i think we can switch directly to ParSeq instead of this👍

Are you maybe interested to open a PR for this model we would be happy to help with !
Otherwise we could take it in the near future we have definitly on track to add more models and i totally agree that ParSeq has to be one of it :)

Do you have experience / tested the models latency on cpu ? Would be interesting to see

[frgfm]

I agree that this could be a good candidate for new text recognition models in 0.6.0 :)
Perhaps we should open a tracker issue for all new models request that we would stage for 0.6.0, or directly link them in the release tracker?

[felixdittrich92]

@frgfm I would say a seperate issue where we can track all requested model additions (splitted in detection / recognition / TF / PT with paper/repo link) and link this issue in the release tracker. wdyt ? Do you like to open it ? :)

[frgfm]

@felixdittrich92 Done :)

[felixdittrich92]

@nikokks @frgfm Do you agree if we close this ticket ? Should be fine if we track it in mentioned issue :)

[frgfm]

I think we should keep this:

• the PR will close this issue, and the release tracker will get the checkbox ticked

So no need to close it, and that will notify @nikokks when this gets resolved, which I guess is of interest to him :)

[felixdittrich92]

👍

[nikokks]

I am inspecting the baudm code on Parseq (https://github.com/baudm/parseq).
I think the code might not be too complicated to integrate.

On my side I managed to connect your ocr_prediction and to integrate the reco_predictor of baudm successfully. The performances are not to good for french documents like yours actually => needs to be finetuned on your secret data 😉

Several questions will come from me on the choices of implementation and integration in doctr:

• can we integrate libraries like timm
• do I have to do it in TF too (torch only if possible)
• if I can add additional utils.py in the future ParSeq folder if this is not your habits.
• etc.

I will clarify my questions rather next weekend 😀

You can close this issue

[felixdittrich92]

Hi @nikokks 👋 ,

lets keep it open for further conversation about ParSeq 👍

About your points:

• timm is a great library but it is mostly pinned in librarys which use it to a fixed version specifier in fact of maintaince and it would be another extra dependency in doctr. So i would suggest to implement ViT from scratch inside the classification models to use it as feature extractor (@frgfm we should move transformers components from models/recognition to models that we can reuse the modules wdyt?)
• it would be great to implement it in both frameworks but it's also fine if you only want to focus on PT ... anyone else could do the translation 👍
• if it is stuff which is only ParSeq related and does not depends on PT or TF i would say base.py is the right place (but lets take a look if we see it in a PR 😅

We are definitly happy to help with. I would say if you are ready open a PR (starting with the classification ViT integration) and we iterate on this wdyt ?

[nikokks]

Hi,

ok for timm.

Other question: can we integrate 'pytorch-lightning~=1.6.5' to the requirements-pt.txt ?

[felixdittrich92]

Hi @nikokks 👋 ,
Why do you think we need this ? :)
We should implement all models in plain pytorch / tensorflow without any wrappers

[nikokks]

ok, it sounds good for me :)

I have added parseq class on my fork.
Now I need to match all args of each methods between your wrapper and the model class parseq =) : (the most difficult part)
I'll do it in the next days or next weekend.

[felixdittrich92]

@nikokks I would suggest the following steps (every should be one PR)

• move models/recognition/transformer to models/modules/transformer (to reuse the implementations we need this for much more models so it should be more global @frgfm wdyt?)
• implement ViT as classification model
• implement the ParSeq Decoder side which use ViT as feature extractor

[frgfm]

• move models/recognition/transformer to models/modules/transformer (to reuse the implementations we need this for much more models so it should be more global @frgfm wdyt?)

I agree 👍

• implement ViT as classification model

Yup, but giving credits to the rightful contributors / source of inspiration when relevant!

[felixdittrich92]

@nikokks Now you can reuse the already implemented transformer parts for ViT 👍

[felixdittrich92]

Hi @nikokks short update i have not forget it will (hopefully) start with ViTSTR next week then it should be easy to implement the decoder from parseq also 👍

[felixdittrich92]

Hi @nikokks,
are you still interested to add ParSeq ? :)
After #1063 ViT has finally found it's way in doctr as backbone.
And #1055 will be updated soon (which could be used as template also for ParSeq)

[nikokks]

Hello, I am currently implementing ParSeq.
It is now working with quite good predictions :)
Do you have any good advice to do a good pull request ?
best

[felixdittrich92]

Hey @nikokks 👋 ,

You can take a look into https://github.com/mindee/doctr/pull/1055/files
which is i think a good template (implementation, tests, docs) for implementing parseq in doctr :)
Otherweise open a PR and we will guide you step by step 👍

PS: If you have only the PT implementation that's fine we can port it later to TF :)

[felixdittrich92]

Finished after #1227 and #1228 are merged :)