app.py

import streamlit as st
import torch
from transformers import RobertaTokenizer, T5ForConditionalGeneration
import pickle
import os
import time
from torch.serialization import safe_globals, add_safe_globals

add_safe_globals([
    "transformers.models.t5.modeling_t5.T5ForConditionalGeneration"
])

# Set page configuration
st.set_page_config(
    page_title="CodeT5 Query Generator",
    page_icon="🤖",
    layout="wide",
)

# CSS styling
st.markdown("""
<style>
    .main-header {
        font-size: 2.5rem;
        color: #4527A0;
        text-align: center;
        margin-bottom: 1rem;
    }
    .sub-header {
        font-size: 1.5rem;
        color: #5E35B1;
        margin-bottom: 1rem;
    }
    .response-container {
        background-color: #f0f2f6;
        border-radius: 10px;
        padding: 20px;
        margin-top: 20px;
    }
    .stButton>button {
        background-color: #673AB7;
        color: white;
    }
    .stButton>button:hover {
        background-color: #5E35B1;
        color: white;
    }
    .footer {
        text-align: center;
        margin-top: 3rem;
        color: #9575CD;
    }
</style>
""", unsafe_allow_html=True)

# App header
st.markdown("<h1 class='main-header'>Network Query Generator</h1>", unsafe_allow_html=True)
st.markdown("<h2 class='sub-header'>Ask questions and get specialized network related queries</h2>", unsafe_allow_html=True)

# Sidebar for model information and settings
with st.sidebar:
    st.title("About")
    st.info("This app uses a fine-tuned CodeT5 model to generate specialized queries from natural language questions.")
    
    st.title("Model Settings")
    max_length = st.slider("Maximum output length", 32, 256, 128)
    num_beams = st.slider("Number of beams", 1, 10, 4)
    temperature = st.slider("Temperature", 0.0, 1.0, 0.7)
    
    st.title("Model Info")
    st.markdown("**Base model:** Salesforce/codet5-small")
    st.markdown("**Fine-tuned on:** Custom dataset")

MODEL_PATH = "finetuned_codet5_small_01.pkl"

# Function to load the model
@st.cache_resource
def load_model(file_path):
    """Load the tokenizer and model using a safe approach"""
    model_name = "Salesforce/codet5-small"
    tokenizer = RobertaTokenizer.from_pretrained(model_name)
    
    try:
        with safe_globals(["transformers.models.t5.modeling_t5.T5ForConditionalGeneration"]):
            model = torch.load(file_path, map_location=torch.device("cuda" if torch.cuda.is_available() else "cpu"), weights_only=False)
        
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model = model.to(device)
        model.eval()
        return tokenizer, model, device, None
    except Exception as e:
        try:
            # Initialize base model
            base_model = T5ForConditionalGeneration.from_pretrained(model_name)
            
            # Load state dict
            with safe_globals(["transformers.models.t5.modeling_t5.T5ForConditionalGeneration"]):
                state_dict = torch.load(file_path, map_location="cpu")
            
            # If the loaded object is already a model, extract just the state dict
            if hasattr(state_dict, 'state_dict'):
                state_dict = state_dict.state_dict()
                
            # Load the state dict into the base model
            base_model.load_state_dict(state_dict)
            
            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
            base_model = base_model.to(device)
            base_model.eval()
            return tokenizer, base_model, device, None
        except Exception as e2:
            return None, None, None, f"Error loading model: {e2}"

# Function to generate query
def generate_query(question, tokenizer, model, device, max_length=128, num_beams=4, temperature=0.7):
    """Generate a query based on the user's question"""
    inputs = tokenizer(question, return_tensors="pt", padding=True, truncation=True, max_length=128)
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        generated_ids = model.generate(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            max_length=max_length,
            num_beams=num_beams,
            temperature=temperature,
            early_stopping=True
        )

    generated_query = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
    return generated_query

# Load the model at startup
with st.spinner("Loading model... (this may take a moment)"):
    tokenizer, model, device, error_message = load_model(MODEL_PATH)
    
    if model is not None:
        st.sidebar.success(f"Model loaded successfully!")
    else:
        st.sidebar.error(f"Failed to load model: {error_message}")

# Main app area
question = st.text_area("Enter your question here:", height=100, placeholder="Example: Show the total data transferred, grouped by user department and VPN type, excluding 'Guest' users, for VPN sessions that lasted longer than 2 hours")

# a button to generate the response
col1, col2, col3 = st.columns([1, 1, 1])
with col2:
    generate_button = st.button("Generate Query", use_container_width=True)

# Display generation result
if generate_button and question:
    if model is not None and tokenizer is not None:
        with st.spinner("Generating response..."):
            # Add a slight delay for user experience
            time.sleep(0.5)
            response = generate_query(
                question, 
                tokenizer, 
                model, 
                device, 
                max_length=max_length, 
                num_beams=num_beams,
                temperature=temperature
            )
        
        st.markdown("<div class='response-container'>", unsafe_allow_html=True)
        st.markdown("### Generated Query:")
        st.code(response, language="sql")
        st.markdown("</div>", unsafe_allow_html=True)
    else:
        st.error("Model could not be loaded. Please check if the model file exists at the correct path.")

# Example questions
with st.expander("Example Questions"):
    example_questions = [
        "Show me the current configuration of the router.",
        "Get the total number of ['firewall', 'router', 'switch', 'server', 'access_point'] with high CPU usage.",
        "Get the total bandwidth usage of access_point FW1.",
        "Get the total uptime of all ['firewall', 'router', 'switch', 'server', 'access_point'].",
        "Get the total number of ['firewall', 'router', 'switch', 'server', 'access_point'].",
        "Find the top 5 devices with the highest data usage, grouped by region and filtering for data usage greater than 10GB."
    ]
    
    for i in range(len(example_questions)):
        st.write(example_questions[i])

# Footer
st.markdown("<p class='footer'>Powered by CodeT5 - Fine-tuned for specialized queries</p>", unsafe_allow_html=True)