app.py

# Import libraries
import json
from pathlib import Path

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel


# Define path
MODEL_DIR = Path("model")
META_PATH = Path("preprocess_meta.json")
THRESH_PATH = MODEL_DIR / "threshold_global.json"

# Load metadata
with META_PATH.open() as f:
    meta = json.load(f)

label2id = meta["label2id"]
id2label = {int(i): label for label, i in label2id.items()}
labels = [id2label[i] for i in range(len(id2label))]
max_length = meta["max_length"]


with THRESH_PATH.open() as f:
    GLOBAL_THRESHOLD = json.load(f)["global_threshold"]

# Load model + LoRA adapters (handles both plain + adapter cases)
tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
base_model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_DIR, trust_remote_code=True
)
try:
    model = PeftModel.from_pretrained(base_model, MODEL_DIR, is_trainable=False)
except ValueError:
    # already merged
    model = base_model

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


def preprocess(subject: str, body: str) -> str:
    sep = tokenizer.sep_token if tokenizer.sep_token else " "
    text = f"{subject.strip()} {sep} {body.strip()}".strip()
    return text

def predict_intents(subject: str, body: str):
    if not subject and not body:
        return {"Predicted labels": [], "Scores": {}}

    text = preprocess(subject, body)
    inputs = tokenizer(
        text,
        truncation=True,
        max_length=max_length,
        return_tensors="pt"
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        logits = model(**inputs).logits.squeeze().float()
        probs = torch.sigmoid(logits).cpu().numpy()

    scores = {label: float(prob) for label, prob in zip(labels, probs)}
    predicted = [label for label, prob in scores.items() if prob >= GLOBAL_THRESHOLD]

    return {
        "Predicted labels": predicted or ["No label ≥ threshold"],
        "Scores": scores,
    }


demo = gr.Interface(
    fn=predict_intents,
    inputs=[
        gr.Textbox(label="Subject", placeholder="Email subject"),
        gr.Textbox(label="Body", lines=12, placeholder="Email body"),
    ],
    outputs=gr.JSON(label="Result (labels & probabilities)"),
    title="Multi-Label Email Intent Classifier",
    description=(
        "DistilBERT + LoRA fine-tuned on synthetic email intents. "
        "Predictions use a global sigmoid threshold of {:.2f}.".format(GLOBAL_THRESHOLD)
    ),
    examples=[
        ["Meeting Reminder: Project Sync", "Dear team, this is a reminder for tomorrow's sync at 10 AM."],
        ["Travel Booking Confirmation", "Your flight to London on 12 June has been confirmed. See attached itinerary."],
    ],
)

if __name__ == "__main__":
    demo.launch()