prosperious's picture
Add 1000 files (ocr-tech/tests/test_audit_limbu_book_missed_line_review_coverage.py ...)
d82db76 verified
|
Raw History Blame Contribute Delete
771 Bytes

PaddleOCR2Pytorch vendored subset

Source: https://github.com/frotms/PaddleOCR2Pytorch

Revision: c05307fbd61e575d55f0fef0334026c7536aec76

License: Apache License 2.0. A copy is included as LICENSE.

Vendored files:

  • modeling/common.py
  • modeling/backbones/rec_lcnetv3.py
  • modeling/backbones/rec_svtrnet.py
  • modeling/necks/rnn.py
  • modeling/heads/rec_ctc_head.py

Local changes are deliberately limited to package-relative imports, MPS-safe tensor placement/contiguity, correct guide detachment, and retaining dynamic recognition width during training. The original fixed-width training pool in PPLCNetV3 was replaced with height-only adaptive pooling so input width remains dynamic and the CTC path produces one time step per four input pixels.