fix: Fixed SAR model for training and inference in PyTorch

[fg-mindee]

Following up on #802, this PR introduces the following changes:

• fixes a major typo in the loss computation (the sequence masking was inverted)
• cleaned encoder and decoder for a much more understandable version

I tried training the model with the synthetic training set, and a real-world validation set and managed to reach almost 100% of exact match after a few epochs.

Any feedback is welcome!

[codecov]

Codecov Report

Merging #831 (58be438) into main (51dc49b) will increase coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #831      +/-   ##
==========================================
+ Coverage   95.97%   95.99%   +0.02%     
==========================================
  Files         131      131              
  Lines        4988     5042      +54     
==========================================
+ Hits         4787     4840      +53     
- Misses        201      202       +1     

Flag	Coverage Δ
unittests	95.99% <100.00%> (+0.02%)	⬆️

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

Impacted Files	Coverage Δ
doctr/models/recognition/sar/pytorch.py	99.20% <100.00%> (+0.06%)	⬆️
doctr/models/detection/linknet/tensorflow.py	97.67% <0.00%> (-1.00%)	⬇️
doctr/models/detection/linknet/pytorch.py	97.97% <0.00%> (-0.89%)	⬇️
doctr/transforms/modules/base.py	94.59% <0.00%> (ø)
doctr/models/classification/resnet/pytorch.py	100.00% <0.00%> (ø)
doctr/models/classification/resnet/tensorflow.py	100.00% <0.00%> (ø)
doctr/models/classification/magc_resnet/pytorch.py	100.00% <0.00%> (ø)
doctr/models/utils/pytorch.py	100.00% <0.00%> (+5.00%)	⬆️

---

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 51dc49b...58be438. Read the comment docs.

[felixdittrich92]

@fg-mindee
Offtopic:
I have faiced the same in my tests (~99% exact) but after loading this trained model in the whole OCR pipeline there was only wrong results 😅 Do you tested this also ? 😄 (train size was in my test cases 1M / val ~ 200k)

[charlesmindee]

Thanks!

[fg-mindee]

@fg-mindee Offtopic: I have faiced the same in my tests (~99% exact) but after loading this trained model in the whole OCR pipeline there was only wrong results sweat_smile Do you tested this also ? smile (train size was in my test cases 1M / val ~ 200k)

@felixdittrich92 I had the same issue until I changed the transformation pipeline during training. On my end, I used a real-world validation dataset, and it worked! If it doesn't on the full OCR pipeline, the problem comes from something else then :)