app.py

"""
app.py — Sentiment Analysis Demo
Deployed on Hugging Face Spaces (Gradio)
NLP Course Project — Dr. Rao Muhammad Adeel Nawab
"""

import gradio as gr
import torch
import re
import os
from transformers import pipeline

# ── Load model — tries your fine-tuned model first, falls back to public ──
HF_TOKEN   = os.environ.get("HF_TOKEN", None)
MY_MODEL   = "Ahmadjunaid/tweet_sentiment_distilbert"
BACKUP     = "distilbert-base-uncased-finetuned-sst-2-english"

try:
    classifier = pipeline(
        "text-classification",
        model=MY_MODEL,
        token=HF_TOKEN,
        device=-1,
    )
    print(f"✅ Loaded custom model: {MY_MODEL}")
except Exception as e:
    print(f"⚠️ Custom model failed ({e}), using backup model.")
    classifier = pipeline(
        "text-classification",
        model=BACKUP,
        device=-1,
    )

# ── Tweet cleaner ──
def clean_tweet(text: str) -> str:
    text = re.sub(r"http\S+|www\S+", "", text)
    text = re.sub(r"@\w+", "", text)
    text = re.sub(r"#(\w+)", r"\1", text)
    text = re.sub(r"[^\w\s!?.,']", " ", text)
    return re.sub(r"\s+", " ", text).strip()

# ── Inference ──
def predict(tweet: str):
    if not tweet.strip():
        return "⚠️ Please enter a tweet.", None, None

    clean  = clean_tweet(tweet)
    result = classifier(clean)[0]
    raw_label = result["label"].upper()
    score     = result["score"]

    if raw_label in ("POSITIVE", "LABEL_1"):
        label, emoji = "Positive", "😊"
        pos_score, neg_score = score, 1 - score
    else:
        label, emoji = "Negative", "😠"
        neg_score, pos_score = score, 1 - score

    result_text = f"{emoji}  {label}   ({score*100:.1f}% confident)"
    bar_data    = {
        "Negative 😠": round(neg_score * 100, 1),
        "Positive 😊": round(pos_score * 100, 1),
    }
    return result_text, bar_data, f'"{clean}"'


EXAMPLES = [
    ["I absolutely love this new phone, it changed my life!"],
    ["This is the worst customer service I have ever experienced."],
    ["Just had coffee. Today feels like a normal day."],
    ["Can't believe how amazing the concert was last night!"],
    ["My flight got cancelled and I missed the meeting. Terrible."],
]

CSS = """
@import url('https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;600;700&display=swap');
body, .gradio-container { background: #0D1B2A !important; font-family: 'Space Grotesk', sans-serif !important; }
#header-block { background: linear-gradient(135deg, #1565C0 0%, #0D47A1 50%, #00ACC1 100%); border-radius: 16px; padding: 32px 40px; margin-bottom: 24px; text-align: center; }
#header-block h1 { font-size: 2.2rem !important; font-weight: 700; color: #fff; margin: 0 0 6px 0; }
#header-block p { color: #B3E5FC; font-size: 0.95rem; margin: 4px 0 0 0; }
label { color: #90CAF9 !important; font-weight: 600 !important; font-size: 0.85rem !important; text-transform: uppercase !important; }
.footer-note { text-align: center; color: #455A6A; font-size: 0.8rem; margin-top: 16px; }
"""

with gr.Blocks(css=CSS, title="Tweet Sentiment Analysis") as demo:
    gr.HTML("""
    <div id="header-block">
        <h1>🐦 Tweet Sentiment Analysis</h1>
        <p>Fine-tuned DistilBERT &middot; Trained on 10,000 English Tweets &middot; F1 = 0.85</p>
        <p style="color:#80DEEA; font-size:0.82rem;">NLP Course Project &middot; Dr. Rao Muhammad Adeel Nawab</p>
    </div>
    """)
    with gr.Row():
        with gr.Column(scale=3):
            tweet_input = gr.Textbox(label="Enter a Tweet", placeholder="e.g. I love this! It completely changed my day 🎉", lines=3)
            with gr.Row():
                submit_btn = gr.Button("Analyse Sentiment →", variant="primary")
                clear_btn  = gr.Button("Clear", variant="secondary")
            gr.Examples(examples=EXAMPLES, inputs=tweet_input, label="Try an Example")
        with gr.Column(scale=2):
            result_out  = gr.Textbox(label="Prediction")
            bar_out     = gr.Json(label="Confidence Scores (%)")
            cleaned_out = gr.Textbox(label="Cleaned Input", lines=2)

    gr.HTML('<div class="footer-note">Model: DistilBERT &middot; Dataset: Sentiment140 &middot; Split: 80/20</div>')
    submit_btn.click(fn=predict, inputs=tweet_input, outputs=[result_out, bar_out, cleaned_out])
    clear_btn.click(lambda: ("", None, ""), outputs=[result_out, bar_out, cleaned_out])
    tweet_input.submit(fn=predict, inputs=tweet_input, outputs=[result_out, bar_out, cleaned_out])

demo.launch()