app.py

import gradio as gr
from indicnlp.transliterate.unicode_transliterate import UnicodeIndicTransliterator
from transformers import VisionEncoderDecoderModel, AutoProcessor, AutoTokenizer
from PIL import Image
import torch
from huggingface_hub import snapshot_download

snapshot_download(repo_id = "QuickHawk/trocr-indic")

ENCODER_MODEL_NAME = "facebook/deit-base-distilled-patch16-224"
DECODER_MODEL_NAME = "ai4bharat/IndicBART"

processor = AutoProcessor.from_pretrained(ENCODER_MODEL_NAME, use_fast=True)
tokenizer = AutoTokenizer.from_pretrained(DECODER_MODEL_NAME, use_fast=True)

model = VisionEncoderDecoderModel.from_pretrained(r"QuickHawk/trocr-indic")

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

LANG_MAP = {
    "as": "Assamese",
    "bn": "Bengali",
    "gu": "Gujarati",
    "hi": "Hindi",
    "kn": "Kannada",
    "ml": "Malayalam",
    "mr": "Marathi",
    "or": "Odia",
    "pa": "Punjabi",
    "ta": "Tamil",
    "te": "Telugu",
    "ur": "Urdu"
}

bos_id = tokenizer._convert_token_to_id_with_added_voc("<s>")
eos_id = tokenizer._convert_token_to_id_with_added_voc("</s>")
pad_id = tokenizer._convert_token_to_id_with_added_voc("<pad>")

def predict(image):

    with torch.no_grad():
        pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)
        outputs_ids = model.generate(
            pixel_values,
            use_cache=True,
            num_beams=4,
            max_length=128,
            min_length=1,
            early_stopping=True,
            pad_token_id=pad_id,
            bos_token_id=bos_id,
            eos_token_id=eos_id,
            decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2en>")
        )

    lang_token = tokenizer.decode(outputs_ids[0][1])
    lang = lang_token[2:-1]

    caption = tokenizer.decode(outputs_ids[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
    return UnicodeIndicTransliterator.transliterate(caption, "hi", lang), LANG_MAP[lang]

gr.Interface(fn=predict, inputs=gr.Image(type="pil"), outputs=[gr.Text(label = "Predicted Text"), gr.Text(label = "Predicted Language")]).launch()