Request access to Lipikar's UnifiedIndicOCR Model

This model is released under CC BY-NC-ND 4.0. Weights are made available for inference and evaluation purposes. Fine-tuning, further training, or creation of derivative models is not permitted under this license.

By requesting access, you confirm that you have read and agree to the CC BY-NC-ND 4.0 license terms, and that you will use this model for non-commercial inference/testing purposes only β€” not for fine-tuning, retraining, or building derivative models.

Log in or Sign Up to review the conditions and access this model content.

UnifiedIndicOCR

Multilingual Indic OCR model based on PARSeq (permuted autoregressive sequence model), with a ViT-L encoder. This repo is self-contained for inference: Only torch, torchvision, timm and Pillow are required.

Repo layout

β”œβ”€β”€ infer.py                 # run this: OCR on one or more images
β”œβ”€β”€ requirements.txt         # torch, torchvision, timm, Pillow β€” nothing else
β”œβ”€β”€ weights/
β”‚   └── parseq_indic.pt      # model weights + architecture/charset config
β”œβ”€β”€ samples/                 # a couple of sample images to try
β”‚   β”œβ”€β”€ adhyaksh_new_1.png
β”‚   └── bn_test_4.jpg
└── strhub/                  # trimmed copy of the original strhub package
                              # just the PARSeq nn.Module + Tokenizer, no
                              # pytorch_lightning/hydra/nltk training code

Model

  • Architecture: PARSeq, ViT-L encoder: embed_dim=1024, enc_depth=24, enc_num_heads=16, 3-layer decoder with dec_num_heads=16. Matches the parseq-L model config combined with configs/10lang.yaml in the training repo (verified against the checkpoint's own saved hyperparameters).
  • Input: 32Γ—128 (HΓ—W) RGB crops, patch_size=[4, 8].
  • Charset: 985 characters (multilingual Indic + Latin + punctuation) + 3 special tokens ([E]/[B]/[P]) = 988 tokens. Note: this is the exact charset stored inside the source checkpoint's own hyperparameters, which has one extra character (ZWJ, U+200D) compared to the checked-in configs/10lang.yaml in the training repo β€” that config file is stale by one character, so the checkpoint's own charset is used here instead of re-deriving it from the yaml.
  • Weights: converted from the training repo's psmr/train-saves/old-a-final-pt.pt (355M params, fp32, ~1.4 GB).

Setup

conda create --name indicocr python=3.10 -y
conda activate indicocr
pip install -r requirements.txt

Run

# Uses the bundled sample images by default
python infer.py

# Or point it at your own images
python infer.py --images path/to/img1.png path/to/img2.jpg --device cuda

Output:

samples/adhyaksh_new_1.png: '...recognized text...'  (confidence: 0.98xx)
samples/bn_test_4.jpg: '...recognized text...'  (confidence: 0.97xx)
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support