app.py

import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForTokenClassification

MODEL_ID = "techysanoj/fine-tuned-IndicNER"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForTokenClassification.from_pretrained(MODEL_ID)

id2label = {int(k): v for k, v in model.config.id2label.items()}

# Color map for Gradio HTML output
COLOR_MAP = {
    "B-PER": "red",
    "I-PER": "red",
    "B-ORG": "green",
    "I-ORG": "green",
    "B-LOC": "blue",
    "I-LOC": "blue",
    "O": "black"
}


def generate_ner_output(text):
    if not text.strip():
        return "Please enter valid input."

    inputs = tokenizer(text, return_tensors="pt")
    token_ids = inputs["input_ids"][0]
    tokens = tokenizer.convert_ids_to_tokens(token_ids)

    with torch.no_grad():
        logits = model(**inputs).logits

    # Softmax for confidence
    probs = F.softmax(logits, dim=-1)[0]

    pred_ids = torch.argmax(probs, dim=-1).tolist()

    html_output = "<div style='font-family: monospace; font-size: 18px;'>"

    for tok, pid, prob_vec in zip(tokens, pred_ids, probs):
        label = id2label[pid]
        conf = float(prob_vec[pid])

        color = COLOR_MAP[label]

        html_output += (
            f"<span style='color:{color}; font-weight:bold;'>"
            f"{tok.replace(' ', '&nbsp;')}</span>"
            f" → <span style='color:{color};'><b>{label}</b></span>"
            f" &nbsp; (conf: {conf:.3f})<br>"
        )

    html_output += "</div>"

    return html_output


# ---------- GRADIO UI -------------
with gr.Blocks() as demo:
    gr.Markdown("## 🔥 IndicNER — Token-Level NER (Colored + Confidence)")

    text_input = gr.Textbox(label="Enter text", lines=3, placeholder="Type sentence here...")
    run_btn = gr.Button("Generate NER")

    ner_html = gr.HTML(label="NER Output")

    run_btn.click(fn=generate_ner_output, inputs=text_input, outputs=ner_html)

demo.launch()