app.py

import gradio as gr
import re
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from nltk.corpus import stopwords
from spaces import GPU  # Required for ZeroGPU Spaces
import nltk

# Download stopwords if not already available
nltk.download("stopwords")
stop_words = set(stopwords.words("english"))

# Define additional words (prepositions, conjunctions, articles) to remove
extra_stopwords = set([
    'a', 'an', 'the', 'and', 'but', 'or', 'for', 'nor', 'so', 'yet', 'at', 'in', 'on', 'with', 'about', 'as', 'by', 'to', 'from', 'of', 'over', 'under', 'during', 'before', 'after', 'between', 'into', 'through', 'among', 'above', 'below'
])

# Combine NLTK stopwords with extra stopwords
stop_words = set(stopwords.words("english")).union(extra_stopwords)


# Model list
model_choices = {
    "DistilBART CNN (sshleifer/distilbart-cnn-12-6)": "sshleifer/distilbart-cnn-12-6",
    "T5 Small (t5-small)": "t5-small",
    "T5 Base (t5-base)": "t5-base",
    "Pegasus XSum (google/pegasus-xsum)": "google/pegasus-xsum",
    "BART CNN (facebook/bart-large-cnn)": "facebook/bart-large-cnn",
}

model_cache = {}

# Clean text: remove special characters and stop words
def clean_text(input_text):
    cleaned = re.sub(r"[^A-Za-z0-9\s]", " ", input_text)
    cleaned_text = re.sub(r"\b[A-Za-z]{2,}[0-9]{3,}\b", "", cleaned_text)  # SKU/product code pattern (letters followed by numbers)

    words = cleaned.split()
    words = [word for word in words if word.lower() not in stop_words]
    return " ".join(words).strip()

# Load model and tokenizer
def load_model(model_name):
    if model_name not in model_cache:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForSeq2SeqLM.from_pretrained(
            model_name, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
        )
        model.to("cuda" if torch.cuda.is_available() else "cpu")
        model_cache[model_name] = (tokenizer, model)

        # Warm up
        dummy_input = tokenizer("summarize: warmup", return_tensors="pt").input_ids.to(model.device)
        model.generate(dummy_input, max_length=10)
    return model_cache[model_name]

# Main function triggered by Gradio
@GPU  # 👈 Required for ZeroGPU to trigger GPU spin-up
def summarize_text(input_text, model_label, char_limit):
    if not input_text.strip():
        return "Please enter some text."

    input_text = clean_text(input_text)
    model_name = model_choices[model_label]
    tokenizer, model = load_model(model_name)

    # Prefix for T5/FLAN-style models
    if "t5" in model_name.lower():
        input_text = "summarize: " + input_text

    inputs = tokenizer(input_text, return_tensors="pt", truncation=True)
    input_ids = inputs["input_ids"].to(model.device)


    # Adjust the generation parameters
    summary_ids = model.generate(
        input_ids,
        max_length=30,                # Keep output length short, around the original text's length
        min_length=15,                # Ensure the summary is not too short
        do_sample=False,              # Disable sampling to avoid introducing new words
        num_beams=5,                  # Beam search to find the most likely sequence of tokens
        early_stopping=True,          # Stop once a reasonable summary is generated
        no_repeat_ngram_size=2        # Prevent repetition of n-grams (bigrams in this case)
    )



    summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
    return summary[:char_limit].strip()

# Gradio UI
iface = gr.Interface(
    fn=summarize_text,
    inputs=[
        gr.Textbox(lines=6, label="Enter text to summarize"),
        gr.Dropdown(choices=list(model_choices.keys()), label="Choose summarization model", value="DistilBART CNN (sshleifer/distilbart-cnn-12-6)"),
        gr.Slider(minimum=30, maximum=200, value=65, step=1, label="Max Character Limit")
    ],
    outputs=gr.Textbox(lines=3, label="Summary (truncated to character limit)"),
    title="🔥 Fast Summarizer (GPU-Optimized)",
    description="Summarizes input using Hugging Face models with ZeroGPU support. Now faster with CUDA, float16, and warm start!"
)

iface.launch()