[datasets][PoC] Enable dataset usage for recognition task

[felixdittrich92]

This PR is handled as Proof of Concept for further discussions to enable the ability to use existing datasets also for recognition task ( main goal: benchmarks ).
It's easier to show the idea directly in code instead of opening a discussion.

Things to investigate if the concept should be fine:

• Synthtext cropping (if use_polygons=True) is to slow (multiprocessing ? maybe in another PR
  reminder: maybe a good reference)

@fg-mindee
No worry we can split this later in parts (maybe geometry, torch, tf) for review if you want 😅

Issue:
#855 First task of this
A good documentation would be part two
(ATTENTION: reminder for docs: SROIE & SVT does only provide uppercase labels and does not match the 'case-sensitive' in images)

Any feedback is very welcome 👍
@charlesmindee @SiddhantBahuguna @fg-mindee

• Part1: [refactor][fix]: Part1 from use datasets for recognition task #889
• Part2: [feature] Part 2 from use datasets for recognition #891

[codecov]

Codecov Report

Merging #867 (31de3ae) into main (9d03085) will increase coverage by 0.11%.
The diff coverage is 98.36%.

@@            Coverage Diff             @@
##             main     #867      +/-   ##
==========================================
+ Coverage   94.82%   94.94%   +0.11%     
==========================================
  Files         133      133              
  Lines        5200     5358     +158     
==========================================
+ Hits         4931     5087     +156     
- Misses        269      271       +2     

Flag	Coverage Δ
unittests	94.94% <98.36%> (+0.11%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
doctr/datasets/utils.py	94.44% <90.00%> (-0.80%)	⬇️
doctr/datasets/synthtext.py	94.73% <92.59%> (-2.41%)	⬇️
doctr/datasets/cord.py	97.72% <100.00%> (+0.29%)	⬆️
doctr/datasets/datasets/pytorch.py	100.00% <100.00%> (ø)
doctr/datasets/datasets/tensorflow.py	100.00% <100.00%> (ø)
doctr/datasets/funsd.py	97.36% <100.00%> (+0.39%)	⬆️
doctr/datasets/ic03.py	97.72% <100.00%> (+0.35%)	⬆️
doctr/datasets/ic13.py	96.77% <100.00%> (+0.62%)	⬆️
doctr/datasets/iiit5k.py	96.96% <100.00%> (+0.19%)	⬆️
doctr/datasets/imgur5k.py	93.33% <100.00%> (+0.83%)	⬆️
... and 18 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9d03085...31de3ae. Read the comment docs.

[felixdittrich92]

@fg-mindee
I saw we have already extract_crops extract_rcrops in models/_utils.py can we maybe move this to geometry ?
provided the PoC is a match 😅

[felixdittrich92]

@frgfm
Same here any feedback ? 😄

[frgfm]

Hey there 🙂

So I had thought about this a few months back. To make sure we are all on the same page, the goal is to:

• take OCR annotated datasets
• make a text recognition dataset out of it

Correct?

If so, two major options arise:

• doing this dynamically (at it getitem)
• doing this statically in the constructor

If this is for training, I'd argue the second option is the one for a few reasons :

• latency
• memory consumption (opening an image 100 times as big to crop it, in a number of samples that isn't necessarily the batch size comes with heavy consequences)

So perhaps this could be done in a temporary or cache folder 🤷‍♂️

What do you think?

[felixdittrich92]

@frgfm
Yes you are right that's the goal (for all currently implemented datasets - without the obj detection one) :)
I have had this ways also in mind so currently i have implemented the 'constructor' option (which works really good) with one problem .... SynthText 😓 it is much to slow and needs a lot of memory (really much ~30gb ram) ...but thanks for your opinon now i can iterate on this and find a way to make it better performing 👍
I was also on track to think about saving for faster reload but it's no option it would waste a lot of the users space
One last thing for the moment are you fine if i move extract_rcrops / extract_crops to geometry otherwise we will have duplicated stuff !? :)
'
EDIT: SynthText also fixed thats the only dataset we need to store as pickle file inside the .cache/doctr/datasets/SynthText so the reloading is also much faster 👍 the other datasets are fine from latency and memory consumption without storing
#889 does the first part
Part2 than datasets + tests

Offtopic: Im really a bit hyped to train the first models on SynthText and MJSynth when we are done with it 😅

[felixdittrich92]

I think it's mostly done i will split it into 2 PRs for easier reviews 👍

[fharper]

Can you fix the issues found by Codacy & CodeFactor please?

[felixdittrich92]

@fharper that' s only a PoC PR will be split into 2 PRs for review 👍
Part 1: #889