fix: Fixed MASTER recognition architecture

[fg-mindee]

This PR introduces the following modifications:

• fixed PyTorch implementation of MAGC module
• refactored MASTER to use magc_resnet31 backbone added to available backbones, and adds pretrained_backbone option
• fixed TF compatibility with AMP for MASTER
• updated unittests accordingly

Any feedback is welcome!

[charlesmindee]

Thanks!

[fg-mindee]

Actually, the paper references the ResNet31 architecture which has 256 channels in stage 2, but in the MASTER paper (Table1) they modified the stage 2 to have 512 channels.

So I reverted back those changes and added back the URL & benchmark perf

[charlesmindee]

Thanks for the revert

[charlesmindee]

Thanks!

[codecov]

Codecov Report

Merging #687 (b90d635) into main (9c45ba6) will decrease coverage by 0.33%.
The diff coverage is 99.23%.

@@            Coverage Diff             @@
##             main     #687      +/-   ##
==========================================
- Coverage   96.40%   96.07%   -0.34%     
==========================================
  Files         117      120       +3     
  Lines        4540     4589      +49     
==========================================
+ Hits         4377     4409      +32     
- Misses        163      180      +17     

Flag	Coverage Δ
unittests	96.07% <99.23%> (-0.34%)	⬇️

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

Impacted Files	Coverage Δ
doctr/models/recognition/crnn/pytorch.py	94.18% <ø> (-5.82%)	⬇️
doctr/models/recognition/transformer/tensorflow.py	88.11% <50.00%> (ø)
doctr/models/backbones/__init__.py	100.00% <100.00%> (ø)
doctr/models/backbones/magc_resnet/__init__.py	100.00% <100.00%> (ø)
doctr/models/backbones/magc_resnet/pytorch.py	100.00% <100.00%> (ø)
doctr/models/backbones/magc_resnet/tensorflow.py	100.00% <100.00%> (ø)
doctr/models/recognition/master/pytorch.py	91.58% <100.00%> (-4.00%)	⬇️
doctr/models/recognition/master/tensorflow.py	95.28% <100.00%> (-1.23%)	⬇️
...dels/detection/differentiable_binarization/base.py	89.44% <0.00%> (-3.34%)	⬇️
... and 4 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 9c45ba6...b90d635. Read the comment docs.