app.py

import streamlit as st
import pandas as pd
import requests
import joblib
import json
import os
import numpy as np
from sklearn.decomposition import PCA

# Load local model components
pca_path = "models/pca.pkl"
local_model_path = "models/model.pkl"

# Load models if they exist

pca = joblib.load(pca_path) if os.path.exists(pca_path) else None
local_model = joblib.load(local_model_path) if os.path.exists(local_model_path) else None

# MLflow Server Endpoint
MLFLOW_SERVER_URL = "http://127.0.0.1:8000/invocations"

# Streamlit UI
st.title("Imago AI Vomitoxin Prediction System")
st.write("Upload a CSV file to predict vomitoxin levels in corn samples.")

# Upload CSV File
uploaded_file = st.file_uploader("Upload CSV File", type=["csv"])

# Prediction Function
def predict(data):
    """Predict using MLflow server if available, else use local model."""
    
    # Try MLflow first
    try:
        st.write("Checking MLflow Server...")
        
        # Convert data to JSON format for MLflow
        json_data = json.dumps({"inputs": df.to_numpy().tolist()})  #  Convert to NumPy first

        
        # Send request to MLflow
        response = requests.post(MLFLOW_SERVER_URL, json={"inputs": json_data})
        
        # If successful, return MLflow prediction
        if response.status_code == 200:
            predictions = response.json()
            st.success("Prediction done using MLflow Server!")
            return np.array(predictions)
        else:
            st.warning("MLflow Server not responding. Falling back to local model.")
    
    except requests.exceptions.RequestException:
        st.warning("MLflow Server is not running. Using local model.")

    # If MLflow fails, use local model
    if local_model and pca:
        # Apply Scaling & PCA
        # scaled_data = scaler.transform(data)
        pca_data = pca.transform(data)
        
        # Predict with local model
        predictions = local_model.predict(pca_data)
        st.success("Prediction done using Local Model!")
        return predictions
    else:
        st.error("No model available for prediction.")
        return None

# Process File & Predict
if uploaded_file:
    try:
        # Read CSV
        df = pd.read_csv(uploaded_file)
        
        # Ensure required columns exist
        if "hsi_id" not in df.columns:
            st.error("CSV file must contain `hsi_id` column.")
        else:
            # Extract Features (Drop hsi_id)
            X = df.drop(columns=["hsi_id"])

            # Make Predictions
            predictions = predict(X)

            if predictions is not None:
                # Create output DataFrame
                df["vomitoxin_ppb"] = predictions
                
                # Mark Harmful Cases (vomitoxin > 1000)
                df["Harmful"] = df["vomitoxin_ppb"].apply(lambda x: "⚠️ Yes" if x > 1000 else "✅ No")
                
                # Display DataFrame
                st.subheader("Prediction Results")
                st.write(df[["hsi_id", "vomitoxin_ppb", "Harmful"]])

                # Highlight Harmful Cases
                st.subheader("Harmful Cases")
                harmful_df = df[df["vomitoxin_ppb"] > 1000]
                st.write(harmful_df[["hsi_id", "vomitoxin_ppb"]])

    except Exception as e:
        st.error(f"Error processing file: {e}")