app.py

import torch
import gradio as gr
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# ----------------------------------------
# 1. Load from Hugging Face Hub
# ----------------------------------------

# Change this to YOUR pushed model repo
HUB_MODEL_ID = "Abelex/amharic-news-bert-multilingual-cased"
  # <--- EDIT IF NEEDED

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MAX_LENGTH = 512  # model context window in TOKENS

# Load tokenizer and model directly from HF Hub
tokenizer = AutoTokenizer.from_pretrained(HUB_MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(HUB_MODEL_ID)
model.to(DEVICE)
model.eval()

# Label mapping from config
id2label = {int(k): v for k, v in model.config.id2label.items()}
num_labels = len(id2label)

# ----------------------------------------
# Helper: highlight tokens after MAX_LENGTH in red (HTML)
# ----------------------------------------
def highlight_token_overflow(text: str, max_tokens: int = 512) -> str:
    """
    Tokenize the input text and generate HTML where tokens beyond
    `max_tokens` are wrapped in red. This shows exactly which tokens
    are outside the model's context window.
    """
    if not text.strip():
        return "<i>No text provided.</i>"

    # Tokenize without truncation (so we can see ALL tokens)
    tokens = tokenizer.tokenize(text)
    if len(tokens) == 0:
        return "<i>No tokens produced by tokenizer.</i>"

    spans = []
    for i, tok in enumerate(tokens):
        # minimal HTML escape
        safe_tok = (
            tok.replace("&", "&amp;")
               .replace("<", "&lt;")
               .replace(">", "&gt;")
        )

        if i >= max_tokens:
            spans.append(f"<span style='color:red;font-weight:bold;'>{safe_tok}</span>")
        else:
            spans.append(f"<span>{safe_tok}</span>")

    html = " ".join(spans)

    if len(tokens) > max_tokens:
        html += (
            f"<br><br>"
            f"<small style='color:red;'>"
            f"Note: Tokens in <b>red</b> are beyond the model context window "
            f"({max_tokens} tokens) and will be truncated."
            f"</small>"
        )
    else:
        html += (
            f"<br><br>"
            f"<small>Token count: {len(tokens)} (≤ {max_tokens}, no truncation).</small>"
        )

    return html

# ----------------------------------------
# 2. Prediction
# ----------------------------------------
def predict_amharic_news(text):
    if not text.strip():
        # Also return highlighted version (empty)
        return "Please enter text.", None, "<i>No text provided.</i>"

    # For actual model inference: truncate to MAX_LENGTH tokens
    encoded = tokenizer(
        text,
        truncation=True,
        padding="max_length",
        max_length=MAX_LENGTH,
        return_tensors="pt"
    )

    encoded = {k: v.to(DEVICE) for k, v in encoded.items()}

    with torch.no_grad():
        outputs = model(**encoded)
        logits = outputs.logits
        probs = torch.softmax(logits, dim=-1).cpu().numpy()[0]

    pred_id = int(np.argmax(probs))
    pred_label = id2label.get(pred_id, f"LABEL_{pred_id}")

    # Prepare probability table
    rows = []
    for i in range(num_labels):
        rows.append((id2label.get(i, f"LABEL_{i}"), float(probs[i])))

    rows = sorted(rows, key=lambda x: x[1], reverse=True)

    # Build HTML showing tokens; tokens >512 in red
    token_highlight_html = highlight_token_overflow(text, max_tokens=MAX_LENGTH)

    # Now we return 3 outputs: prediction, probs table, token visualization
    return f"Predicted Label: {pred_label}", rows, token_highlight_html

# ----------------------------------------
# 3. Gradio Interface
# ----------------------------------------
demo = gr.Interface(
    fn=predict_amharic_news,
    inputs=gr.Textbox(
        lines=5,
        label="Enter Amharic News Text",
        placeholder="እባክዎ የአማርኛ ዜና ጽሑፍ ያስገቡ..."
    ),
    outputs=[
        gr.Textbox(label="Prediction"),
        gr.Dataframe(
            headers=["Label", "Probability"],
            label="Class Probabilities"
        ),
        gr.HTML(label="Tokenizer view (tokens > 512 are red)")
    ],
    title="Amharic News Classifier",
    description=(
        "XLM-RoBERTa model loaded directly from Hugging Face Hub (raw text input, no preprocessing). "
        "Below, tokenizer output shows which tokens are beyond the 512-token context window (in red)."
    )
)

demo.launch()