Update src/streamlit_app.py

src/streamlit_app.py

@@ -1,653 +1,572 @@
 """
-Streamlit Dashboard for DLRM Book Recommendation System
-Simple interface for DLRM-based book recommendations
+DLRM Inference Engine for Book Recommendations
+Loads trained DLRM model and provides recommendation functionality
 """
 
 import os
 import sys
-import streamlit as st
-
-# Check if CPU_ONLY mode is enabled via command line argument
-if len(sys.argv) > 1 and sys.argv[1] == '--cpu-only':
-    os.environ['CPU_ONLY'] = 'true'
-    print("🔄 Running in CPU-only mode (CUDA disabled)")
-
-import pandas as pd
-import numpy as np
 import torch
+import numpy as np
+import pandas as pd
 import pickle
-from typing import Dict, List, Tuple, Optional
+import mlflow
+from mlflow import MlflowClient
+import tempfile
+from typing import List, Dict, Tuple, Optional, Any
+from functools import partial
 import warnings
 warnings.filterwarnings('ignore')
 
-# Import our DLRM recommender
+# Check for CPU_ONLY environment variable
+CPU_ONLY = os.environ.get('CPU_ONLY', 'false').lower() == 'true'
+
+# Disable CUDA if CPU_ONLY is set
+if CPU_ONLY:
+    os.environ['CUDA_VISIBLE_DEVICES'] = ''
+    print("🔄 Running in CPU-only mode (CUDA disabled)")
+
 try:
-    from dlrm_inference import DLRMBookRecommender, load_dlrm_recommender, TORCHREC_AVAILABLE
+    from torchrec import EmbeddingBagCollection
+    from torchrec.models.dlrm import DLRM, DLRMTrain
+    from torchrec.modules.embedding_configs import EmbeddingBagConfig
+    from torchrec.sparse.jagged_tensor import KeyedJaggedTensor
+    from torchrec.datasets.utils import Batch
+    TORCHREC_AVAILABLE = True
 except ImportError as e:
-    print(f"⚠️ Error importing DLRM recommender: {e}")
+    print(f"⚠️ Warning: torchrec import error: {e}")
+    print("⚠️ Some functionality will be limited")
     TORCHREC_AVAILABLE = False
 
-
-# Page configuration
-st.set_page_config(
-    page_title="DLRM Book Recommendations",
-    page_icon="📚",
-    layout="wide",
-    initial_sidebar_state="expanded"
-)
-
-# Check if running in CPU-only mode
-cpu_only_mode = os.environ.get('CPU_ONLY', 'false').lower() == 'true'
-
-# Custom CSS
-st.markdown("""
-<style>
-    .main-header {
-        font-size: 3rem;
-        color: #1f77b4;
-        text-align: center;
-        margin-bottom: 2rem;
-    }
-    .metric-card {
-        background-color: #f0f2f6;
-        padding: 1rem;
-        border-radius: 0.5rem;
-        border-left: 5px solid #1f77b4;
-    }
-    .dlrm-explanation {
-        background-color: #e8f4fd;
-        padding: 1rem;
-        border-radius: 0.5rem;
-        border-left: 4px solid #0066cc;
-        margin: 1rem 0;
-    }
-    .book-card {
-        background-color: #ffffff;
-        padding: 1rem;
-        border-radius: 0.5rem;
-        border: 1px solid #e1e5eb;
-        margin-bottom: 1rem;
-    }
-    .cpu-mode-banner {
-        background-color: #fff3cd;
-        color: #856404;
-        padding: 0.75rem;
-        border-radius: 0.5rem;
-        border-left: 4px solid #ffeeba;
-        margin: 1rem 0;
-        text-align: center;
-    }
-</style>
-""", unsafe_allow_html=True)
-
-@st.cache_data
-def load_data():
-    """Load and cache the book data"""
-    try:
-        books_df = pd.read_csv('Books.csv', encoding='latin-1', low_memory=False)
-        users_df = pd.read_csv('Users.csv', encoding='latin-1', low_memory=False)
-        ratings_df = pd.read_csv('Ratings.csv', encoding='latin-1', low_memory=False)
-        
-        # Clean column names
-        books_df.columns = books_df.columns.str.replace('"', '')
-        users_df.columns = users_df.columns.str.replace('"', '')
-        ratings_df.columns = ratings_df.columns.str.replace('"', '')
-        
-        return books_df, users_df, ratings_df
-    except Exception as e:
-        st.error(f"Error loading data: {e}")
-        return None, None, None
-
-@st.cache_resource
-def load_dlrm_model():
-    """Load and cache the DLRM model"""
-
+class DLRMBookRecommender:
+    """DLRM-based book recommender for inference"""
     
-    try:
-        recommender = load_dlrm_recommender("file")
-        return recommender
-    except Exception as e:
-        st.error(f"Error loading DLRM model: {e}")
-        return None
-
-def display_book_info(book_isbn, books_df, show_rating=None):
-    """Display book information with actual book cover"""
-    book_info = books_df[books_df['ISBN'] == book_isbn]
-    
-    if len(book_info) == 0:
-        st.write(f"Book with ISBN {book_isbn} not found")
-        return
-    
-    book = book_info.iloc[0]
-    
-    col1, col2 = st.columns([1, 3])
+    def __init__(self, model_path: str = None, run_id: str = None):
+        """
+        Initialize DLRM book recommender
+        
+        Args:
+            model_path: Path to saved model state dict
+            run_id: MLflow run ID to load model from
+        """
+        self.device = torch.device("cpu")
+        self.model = None
+        self.preprocessing_info = None
+        self.torchrec_available = TORCHREC_AVAILABLE
+        
+        if not self.torchrec_available:
+            print("⚠️ Running in limited mode without torchrec")
+            return
+            
+        # Load preprocessing info
+        self._load_preprocessing_info()
+        
+        # Load model
+        if model_path and os.path.exists(model_path):
+            self._load_model_from_path(model_path)
+        elif run_id:
+            self._load_model_from_mlflow(run_id)
+        else:
+            print("⚠️ No model loaded. Please provide model_path or run_id")
     
-    with col1:
-        # Try to display actual book cover from Image-URL-M
-        image_url = book.get('Image-URL-M', '')
-        
-        if image_url and pd.notna(image_url) and str(image_url) != 'nan':
-            try:
-                # Clean the URL (sometimes there are issues with Amazon URLs)
-                clean_url = str(image_url).strip()
-                if clean_url and 'http' in clean_url:
-                    st.image(clean_url, width=150, caption="📚")
-                else:
-                    # Fallback to placeholder
-                    st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width=150)
-            except Exception as e:
-                # If image loading fails, show placeholder
-                st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width=150)
-                st.caption("⚠️ Cover unavailable")
+    def _load_preprocessing_info(self):
+        """Load preprocessing information"""
+        if os.path.exists('book_dlrm_preprocessing.pkl'):
+            with open('book_dlrm_preprocessing.pkl', 'rb') as f:
+                self.preprocessing_info = pickle.load(f)
+            
+            self.dense_cols = self.preprocessing_info['dense_cols']
+            self.cat_cols = self.preprocessing_info['cat_cols']
+            self.emb_counts = self.preprocessing_info['emb_counts']
+            self.user_encoder = self.preprocessing_info['user_encoder']
+            self.book_encoder = self.preprocessing_info['book_encoder']
+            self.publisher_encoder = self.preprocessing_info['publisher_encoder']
+            self.location_encoder = self.preprocessing_info['location_encoder']
+            self.scaler = self.preprocessing_info['scaler']
+            
+            print("✅ Preprocessing info loaded")
         else:
-            # Show placeholder if no image URL
-            st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width=150)
-            st.caption("📚 No cover")
+            raise FileNotFoundError("book_dlrm_preprocessing.pkl not found. Run preprocessing first.")
     
-    with col2:
-        st.markdown(f"**{book['Book-Title']}**")
-        st.write(f"*by {book['Book-Author']}*")
-        st.write(f"📅 Published: {book.get('Year-Of-Publication', 'Unknown')}")
-        st.write(f"🏢 Publisher: {book.get('Publisher', 'Unknown')}")
-        st.write(f"📖 ISBN: {book['ISBN']}")
-        
-        if show_rating is not None:
-            st.markdown(f"**🎯 DLRM Score: {show_rating:.4f}**")
+    def _load_model_from_path(self, model_path: str):
+        """Load model from saved state dict"""
+        try:
+            # Create model architecture
+            eb_configs = [
+                EmbeddingBagConfig(
+                    name=f"t_{feature_name}",
+                    embedding_dim=64,  # Default embedding dim
+                    num_embeddings=self.emb_counts[feature_idx],
+                    feature_names=[feature_name],
+                )
+                for feature_idx, feature_name in enumerate(self.cat_cols)
+            ]
 
-def main():
-    # Header
-    st.markdown('<h1 class="main-header">📚 DLRM Book Recommendation System</h1>', unsafe_allow_html=True)
-    st.markdown("### Deep Learning Recommendation Model for Personalized Book Suggestions")
-    
-    # CPU Mode Banner (if enabled)
-    if cpu_only_mode:
-        st.markdown('<div class="cpu-mode-banner">⚙️ Running in CPU-only mode (NVIDIA drivers not required)</div>', unsafe_allow_html=True)
-    
-    st.markdown("---")
+            dlrm_model = DLRM(
+                embedding_bag_collection=EmbeddingBagCollection(
+                    tables=eb_configs, device=self.device
+                ),
+                dense_in_features=len(self.dense_cols),
+                dense_arch_layer_sizes=[256, 128, 64],
+                over_arch_layer_sizes=[512, 256, 128, 1],
+                dense_device=self.device,
+            )
 
+            # Load state dict
+            state_dict = torch.load(model_path, map_location=self.device)
+            
+            # Remove 'model.' prefix if present
+            if any(key.startswith('model.') for key in state_dict.keys()):
+                state_dict = {k[6:]: v for k, v in state_dict.items()}
+            
+            dlrm_model.load_state_dict(state_dict)
+            self.model = dlrm_model
+            self.model.eval()
+            
+            print(f"✅ Model loaded from {model_path}")
+            
+        except Exception as e:
+            print(f"❌ Error loading model: {e}")
     
-    # Load data
-    with st.spinner("Loading book data..."):
-        books_df, users_df, ratings_df = load_data()
-    
-    if books_df is None:
-        st.error("Failed to load data. Please check if CSV files are available.")
-        return
-    
-    # Sidebar info
-    st.sidebar.title("📊 Dataset Information")
-    st.sidebar.metric("📚 Books", f"{len(books_df):,}")
-    st.sidebar.metric("👥 Users", f"{len(users_df):,}")
-    st.sidebar.metric("⭐ Ratings", f"{len(ratings_df):,}")
-    
-    # Load DLRM model
-    with st.spinner("Loading DLRM model..."):
-        recommender = load_dlrm_model()
-    
-    if recommender is None or not hasattr(recommender, 'model') or recommender.model is None:
-        if cpu_only_mode:
-            st.warning("⚠️ DLRM model not available in CPU-only mode")
-            st.info("The app will continue with limited functionality")
+    def _load_model_from_mlflow(self, run_id: str):
+        """Load model from MLflow"""
+        try:
+            client = MlflowClient()
+            run = client.get_run(run_id)
+            
+            # Get model parameters from MLflow
+            params = run.data.params
+            cat_cols = eval(params.get('cat_cols'))
+            emb_counts = eval(params.get('emb_counts'))
+            dense_cols = eval(params.get('dense_cols'))
+            embedding_dim = int(params.get('embedding_dim', 64))
+            dense_arch_layer_sizes = eval(params.get('dense_arch_layer_sizes'))
+            over_arch_layer_sizes = eval(params.get('over_arch_layer_sizes'))
             
-            # Show options for browsing books without recommendations
-            st.subheader("📚 Browse Books")
+            # Download model from MLflow
+            temp_dir = tempfile.mkdtemp()
             
-            # Simple book browser
-            search_query = st.text_input("Search for books", placeholder="Enter title, author, or publisher")
-            if search_query:
-                mask = (
-                    books_df['Book-Title'].str.contains(search_query, case=False, na=False) |
-                    books_df['Book-Author'].str.contains(search_query, case=False, na=False) |
-                    books_df['Publisher'].str.contains(search_query, case=False, na=False)
+            # Try different artifact paths
+            for artifact_path in ['model_state_dict_final', 'model_state_dict_2', 'model_state_dict_1', 'model_state_dict_0']:
+                try:
+                    client.download_artifacts(run_id, f"{artifact_path}/state_dict.pth", temp_dir)
+                    state_dict = mlflow.pytorch.load_state_dict(f"{temp_dir}/{artifact_path}")
+                    break
+                except:
+                    continue
+            else:
+                raise Exception("No model artifacts found")
+            
+            # Create model
+            eb_configs = [
+                EmbeddingBagConfig(
+                    name=f"t_{feature_name}",
+                    embedding_dim=embedding_dim,
+                    num_embeddings=emb_counts[feature_idx],
+                    feature_names=[feature_name],
                 )
-                results = books_df[mask].head(20)
-                
-                if len(results) > 0:
-                    st.success(f"Found {len(results)} books matching '{search_query}'")
-                    for _, book in results.iterrows():
-                        st.markdown(f"**{book['Book-Title']}** by *{book['Book-Author']}*")
-                        st.write(f"Published: {book.get('Year-Of-Publication', 'Unknown')} | ISBN: {book['ISBN']}")
-                        st.markdown("---")
-                else:
-                    st.info(f"No books found matching '{search_query}'")
+                for feature_idx, feature_name in enumerate(cat_cols)
+            ]
+
+            dlrm_model = DLRM(
+                embedding_bag_collection=EmbeddingBagCollection(
+                    tables=eb_configs, device=self.device
+                ),
+                dense_in_features=len(dense_cols),
+                dense_arch_layer_sizes=dense_arch_layer_sizes,
+                over_arch_layer_sizes=over_arch_layer_sizes,
+                dense_device=self.device,
+            )
+
+            # Remove prefix and load state dict
+            if any(key.startswith('model.') for key in state_dict.keys()):
+                state_dict = {k[6:]: v for k, v in state_dict.items()}
             
-            return
-        else:
-            st.error("❌ DLRM model not available")
-            st.info("Please run the training script first: `python train_dlrm_books.py`")
+            dlrm_model.load_state_dict(state_dict)
+            self.model = dlrm_model
+            self.model.eval()
             
-            st.markdown("### Available Options:")
-            st.markdown("1. **Train DLRM Model**: Run `python train_dlrm_books.py`")
-            st.markdown("2. **Prepare Data**: Run `python dlrm_book_recommender.py`")
-            st.markdown("3. **Check Files**: Ensure preprocessing files exist")
-            st.markdown("4. **Try CPU-only Mode**: Run `streamlit run streamlit_dlrm_app.py -- --cpu-only`")
+            print(f"✅ Model loaded from MLflow run: {run_id}")
             
-            return
-    
-    if cpu_only_mode:
-        st.success("✅ DLRM model loaded successfully in CPU-only mode!")
-    else:
-        st.success("✅ DLRM model loaded successfully!")
-    
-    # Model info
-    st.sidebar.markdown("---")
-    st.sidebar.subheader("🤖 DLRM Model Info")
-    if recommender.preprocessing_info:
-        st.sidebar.write(f"Dense features: {len(recommender.dense_cols)}")
-        st.sidebar.write(f"Categorical features: {len(recommender.cat_cols)}")
-        st.sidebar.write(f"Embedding dim: 64")
-    
-    # Main interface
-    tab1, tab2, tab3, tab4 = st.tabs(["🎯 Get Recommendations", "🔍 Test Predictions", "📊 Model Analysis", "📸 Book Gallery"])
+        except Exception as e:
+            print(f"❌ Error loading model from MLflow: {e}")
     
-    with tab1:
-        st.header("🎯 DLRM Book Recommendations")
-        st.info("Get personalized book recommendations using the trained DLRM model")
-        
-        # User selection
-        col1, col2 = st.columns([2, 1])
-        
-        with col1:
-            user_ids = sorted(users_df['User-ID'].unique())
-            selected_user_id = st.selectbox("Select a user", user_ids[:1000])  # Limit for performance
-        
-        with col2:
-            num_recommendations = st.slider("Number of recommendations", 5, 20, 10)
-        
-        # Show user info
-        user_info = users_df[users_df['User-ID'] == selected_user_id]
-        if len(user_info) > 0:
-            user = user_info.iloc[0]
-            st.markdown(f"**User Info**: Age: {user.get('Age', 'Unknown')}, Location: {user.get('Location', 'Unknown')}")
-        
-        # User's reading history
-        user_ratings = ratings_df[ratings_df['User-ID'] == selected_user_id]
-        if len(user_ratings) > 0:
-            with st.expander(f"📖 User's Reading History ({len(user_ratings)} books)", expanded=False):
-                top_rated = user_ratings.sort_values('Book-Rating', ascending=False).head(10)
-                for _, rating in top_rated.iterrows():
-                    book_info = books_df[books_df['ISBN'] == rating['ISBN']]
-                    if len(book_info) > 0:
-                        book = book_info.iloc[0]
-                        st.write(f"• **{book['Book-Title']}** by {book['Book-Author']} - {rating['Book-Rating']}/10 ⭐")
-        
-        if st.button("🚀 Get DLRM Recommendations", type="primary"):
-            with st.spinner("🤖 DLRM is analyzing user preferences..."):
-                
-                # Get candidate books (popular books not rated by user)
-                user_rated_books = set(user_ratings['ISBN']) if len(user_ratings) > 0 else set()
-                
-                # Get popular books as candidates
-                book_popularity = ratings_df.groupby('ISBN').size().sort_values(ascending=False)
-                candidate_books = [isbn for isbn in book_popularity.head(100).index if isbn not in user_rated_books]
-                
-                if len(candidate_books) < num_recommendations:
-                    candidate_books = book_popularity.head(200).index.tolist()
-                
-                # Get recommendations
-                recommendations = recommender.get_user_recommendations(
-                    user_id=selected_user_id,
-                    candidate_books=candidate_books,
-                    k=num_recommendations
-                )
-            
-            if recommendations:
-                st.success(f"Generated {len(recommendations)} DLRM recommendations!")
-                
-                st.subheader("🎯 DLRM Recommendations")
-                
-                for i, (book_isbn, score) in enumerate(recommendations, 1):
-                    book_info = books_df[books_df['ISBN'] == book_isbn]
-                    if len(book_info) > 0:
-                        with st.expander(f"{i}. Recommendation (DLRM Score: {score:.4f})", expanded=(i <= 3)):
-                            display_book_info(book_isbn, books_df, show_rating=score)
-                            
-                            # Additional book stats
-                            book_ratings = ratings_df[ratings_df['ISBN'] == book_isbn]
-                            if len(book_ratings) > 0:
-                                avg_rating = book_ratings['Book-Rating'].mean()
-                                num_ratings = len(book_ratings)
-                                
-                                st.markdown('<div class="dlrm-explanation">', unsafe_allow_html=True)
-                                st.markdown("**📊 Book Statistics:**")
-                                st.write(f"Average Rating: {avg_rating:.1f}/10 from {num_ratings} readers")
-                                st.write(f"DLRM Confidence: {score:.1%}")
-                                st.markdown('</div>', unsafe_allow_html=True)
-                    else:
-                        st.write(f"Book with ISBN {book_isbn} not found in database")
-            else:
-                st.warning("No recommendations generated")
+    def _prepare_user_features(self, user_id: int, user_data: Optional[Dict] = None) -> Tuple[torch.Tensor, KeyedJaggedTensor]:
+        """Prepare user features for inference"""
+        
+        if user_data is None:
+            # Create default user features
+            user_data = {
+                'User-ID': user_id,
+                'Age': 30,  # Default age
+                'Location': 'usa',  # Default location
+            }
+        
+        # Encode categorical features
+        try:
+            user_id_encoded = self.user_encoder.transform([str(user_id)])[0]
+        except:
+            # Handle unknown user
+            user_id_encoded = 0
+        
+        try:
+            location = str(user_data.get('Location', 'usa')).split(',')[-1].strip().lower()
+            country_encoded = self.location_encoder.transform([location])[0]
+        except:
+            country_encoded = 0
+        
+        # Age group
+        age = user_data.get('Age', 30)
+        if age < 18:
+            age_group = 0
+        elif age < 25:
+            age_group = 1
+        elif age < 35:
+            age_group = 2
+        elif age < 50:
+            age_group = 3
+        elif age < 65:
+            age_group = 4
+        else:
+            age_group = 5
+        
+        # Get user statistics (if available)
+        user_activity = user_data.get('user_activity', 10)  # Default
+        user_avg_rating = user_data.get('user_avg_rating', 6.0)  # Default
+        age_normalized = user_data.get('Age', 30)
+        
+        # Normalize dense features
+        dense_features = np.array([[age_normalized, 2000, user_activity, 10, user_avg_rating, 6.0]])  # Default values
+        dense_features = self.scaler.transform(dense_features)
+        dense_features = torch.tensor(dense_features, dtype=torch.float32)
+        
+        return dense_features, user_id_encoded, country_encoded, age_group
     
-    with tab2:
-        st.header("🔍 Test DLRM Predictions")
-        st.info("Test how well DLRM predicts actual user ratings")
-        
-        col1, col2 = st.columns(2)
-        
-        with col1:
-            test_user_id = st.selectbox("Select user for testing", user_ids[:500], key="test_user")
-        
-        with col2:
-            test_mode = st.radio("Test mode", ["Random books", "User's actual books"])
-        
-        if st.button("🧪 Test Predictions", type="secondary"):
-            with st.spinner("Testing DLRM predictions..."):
-                
-                if test_mode == "User's actual books":
-                    # Test on user's actual rated books
-                    user_test_ratings = ratings_df[ratings_df['User-ID'] == test_user_id].sample(min(10, len(user_ratings)))
-                    
-                    if len(user_test_ratings) > 0:
-                        st.subheader("🎯 DLRM vs Actual Ratings")
-                        
-                        predictions = []
-                        actuals = []
-                        
-                        for _, rating in user_test_ratings.iterrows():
-                            book_isbn = rating['ISBN']
-                            actual_rating = rating['Book-Rating']
-                            
-                            # Get DLRM prediction
-                            dlrm_score = recommender.predict_rating(test_user_id, book_isbn)
-                            
-                            predictions.append(dlrm_score)
-                            actuals.append(actual_rating >= 6)  # Convert to binary
-                            
-                            # Display comparison
-                            book_info = books_df[books_df['ISBN'] == book_isbn]
-                            if len(book_info) > 0:
-                                book = book_info.iloc[0]
-                                
-                                col1, col2, col3 = st.columns([2, 1, 1])
-                                with col1:
-                                    st.write(f"**{book['Book-Title']}**")
-                                    st.write(f"*by {book['Book-Author']}*")
-                                
-                                with col2:
-                                    st.metric("Actual Rating", f"{actual_rating}/10")
-                                
-                                with col3:
-                                    st.metric("DLRM Score", f"{dlrm_score:.3f}")
-                        
-                        # Calculate accuracy
-                        if predictions and actuals:
-                            # Convert DLRM scores to binary predictions
-                            binary_preds = [1 if p > 0.5 else 0 for p in predictions]
-                            accuracy = sum(p == a for p, a in zip(binary_preds, actuals)) / len(actuals)
-                            
-                            st.markdown("---")
-                            st.success(f"🎯 DLRM Accuracy: {accuracy:.1%}")
-                            
-                            # Show correlation
-                            actual_numeric = [rating['Book-Rating'] for _, rating in user_test_ratings.iterrows()]
-                            correlation = np.corrcoef(predictions, actual_numeric)[0, 1] if len(predictions) > 1 else 0
-                            st.info(f"📊 Correlation with actual ratings: {correlation:.3f}")
-                    
-                    else:
-                        st.warning("No ratings found for this user")
-                
-                else:
-                    # Test on random books
-                    random_books = books_df.sample(10)['ISBN'].tolist()
-                    
-                    st.subheader("🎲 Random Book Predictions")
-                    
-                    for book_isbn in random_books:
-                        dlrm_score = recommender.predict_rating(test_user_id, book_isbn)
-                        
-                        book_info = books_df[books_df['ISBN'] == book_isbn]
-                        if len(book_info) > 0:
-                            book = book_info.iloc[0]
-                            
-                            col1, col2 = st.columns([3, 1])
-                            with col1:
-                                st.write(f"**{book['Book-Title']}** by *{book['Book-Author']}*")
-                            
-                            with col2:
-                                st.metric("DLRM Score", f"{dlrm_score:.4f}")
+    def _prepare_book_features(self, book_isbn: str, book_data: Optional[Dict] = None) -> Tuple[int, int, int, int]:
+        """Prepare book features for inference"""
+        
+        if book_data is None:
+            book_data = {}
+        
+        # Encode book ID
+        try:
+            book_id_encoded = self.book_encoder.transform([str(book_isbn)])[0]
+        except:
+            book_id_encoded = 0
+        
+        # Encode publisher
+        try:
+            publisher = str(book_data.get('Publisher', 'Unknown'))
+            publisher_encoded = self.publisher_encoder.transform([publisher])[0]
+        except:
+            publisher_encoded = 0
+        
+        # Publication decade
+        year = book_data.get('Year-Of-Publication', 2000)
+        decade = ((int(year) // 10) * 10)
+        try:
+            decade_encoded = preprocessing_info.get('decade_encoder', LabelEncoder()).transform([str(decade)])[0]
+        except:
+            decade_encoded = 6  # Default to 2000s
+        
+        # Rating level (default to medium)
+        rating_level = 1
+        
+        return book_id_encoded, publisher_encoded, decade_encoded, rating_level
     
-    with tab3:
-        st.header("📊 DLRM Model Analysis")
-        st.info("Analysis of the DLRM model performance and characteristics")
+    def predict_rating(self, user_id: int, book_isbn: str, 
+                      user_data: Optional[Dict] = None, 
+                      book_data: Optional[Dict] = None) -> float:
+        """
+        Predict rating probability for user-book pair
         
-        # Model architecture info
-        if recommender and recommender.preprocessing_info:
-            col1, col2 = st.columns(2)
+        Args:
+            user_id: User ID
+            book_isbn: Book ISBN
+            user_data: Additional user data (optional)
+            book_data: Additional book data (optional)
             
-            with col1:
-                st.subheader("🏗️ Model Architecture")
-                st.write(f"**Dense Features ({len(recommender.dense_cols)}):**")
-                for col in recommender.dense_cols:
-                    st.write(f"• {col}")
-                
-                st.write(f"**Categorical Features ({len(recommender.cat_cols)}):**")
-                for i, col in enumerate(recommender.cat_cols):
-                    st.write(f"• {col}: {recommender.emb_counts[i]} embeddings")
+        Returns:
+            Prediction probability (0-1)
+        """
+        if self.model is None:
+            print("❌ Model not loaded")
+            return 0.0
+        
+        if not self.torchrec_available:
+            print("❌ Cannot make predictions without torchrec")
+            return 0.5  # Return default neutral prediction
+        
+        try:
+            # Prepare features
+            dense_features, user_id_encoded, country_encoded, age_group = self._prepare_user_features(user_id, user_data)
+            book_id_encoded, publisher_encoded, decade_encoded, rating_level = self._prepare_book_features(book_isbn, book_data)
             
-            with col2:
-                st.subheader("📈 Dataset Statistics")
-                total_samples = recommender.preprocessing_info.get('total_samples', 0)
-                positive_rate = recommender.preprocessing_info.get('positive_rate', 0)
-                
-                st.metric("Total Samples", f"{total_samples:,}")
-                st.metric("Positive Rate", f"{positive_rate:.1%}")
-                st.metric("Train Samples", f"{recommender.preprocessing_info.get('train_samples', 0):,}")
-                st.metric("Validation Samples", f"{recommender.preprocessing_info.get('val_samples', 0):,}")
-                st.metric("Test Samples", f"{recommender.preprocessing_info.get('test_samples', 0):,}")
-        
-        # Feature importance analysis
-        st.subheader("🔍 Feature Analysis")
-        
-        if st.button("Analyze Feature Importance"):
-            with st.spinner("Analyzing feature importance..."):
-                
-                # Sample some users and books
-                sample_users = users_df['User-ID'].sample(20).tolist()
-                sample_books = books_df['ISBN'].sample(20).tolist()
-                
-                # Test different feature combinations
-                st.write("**Feature Impact Analysis:**")
-                
-                base_predictions = []
-                for user_id in sample_users[:5]:
-                    for book_isbn in sample_books[:5]:
-                        score = recommender.predict_rating(user_id, book_isbn)
-                        base_predictions.append(score)
-                
-                avg_prediction = np.mean(base_predictions)
-                st.metric("Average Prediction Score", f"{avg_prediction:.4f}")
-                
-                st.success("✅ Feature analysis completed!")
-        
-        # Load training results if available
-        if os.path.exists('dlrm_book_training_results.pkl'):
-            with open('/home/mr-behdadi/PROJECT/ICE/dlrm_book_training_results.pkl', 'rb') as f:
-                training_results = pickle.load(f)
+            # Create sparse features
+            kjt_values = [user_id_encoded, book_id_encoded, publisher_encoded, country_encoded, age_group, decade_encoded, rating_level]
+            kjt_lengths = [1] * len(kjt_values)
             
-            st.subheader("📈 Training Results")
+            sparse_features = KeyedJaggedTensor.from_lengths_sync(
+                self.cat_cols,
+                torch.tensor(kjt_values),
+                torch.tensor(kjt_lengths, dtype=torch.int32),
+            )
             
-            col1, col2 = st.columns(2)
+            # Make prediction
+            with torch.no_grad():
+                logits = self.model(dense_features=dense_features, sparse_features=sparse_features)
+                prediction = torch.sigmoid(logits).item()
             
-            with col1:
-                st.metric("Final Validation AUROC", f"{training_results.get('final_val_auroc', 0):.4f}")
-                st.metric("Test AUROC", f"{training_results.get('test_auroc', 0):.4f}")
+            return prediction
             
-            with col2:
-                val_history = training_results.get('val_aurocs_history', [])
-                if val_history:
-                    st.line_chart(pd.DataFrame({
-                        'Epoch': range(len(val_history)),
-                        'Validation AUROC': val_history
-                    }).set_index('Epoch'))
-    
-    # Instructions
-    st.markdown("---")
-    st.markdown("""
-    ## 🚀 How DLRM Works for Book Recommendations
-    
-    **DLRM (Deep Learning Recommendation Model)** is specifically designed for recommendation systems and offers several advantages:
-    
-    ### 🏗️ Architecture Benefits:
-    - **Multi-feature Processing**: Handles both categorical (user ID, book ID, publisher) and numerical (age, ratings) features
-    - **Embedding Tables**: Learns rich representations for categorical features
-    - **Cross-feature Interactions**: Captures complex relationships between different features
-    - **Scalable Design**: Efficiently handles large-scale recommendation datasets
+        except Exception as e:
+            print(f"Error in prediction: {e}")
+            return 0.0
     
-    ### 📊 Features Used:
-    **Categorical Features:**
-    - User ID, Book ID, Publisher, Country, Age Group, Publication Decade, Rating Level
-    
-    **Dense Features:**  
-    - Normalized Age, Publication Year, User Activity, Book Popularity, Average Ratings
-    
-    ### 🎯 Why DLRM vs LLM for Recommendations:
-    - **Purpose-built**: Specifically designed for recommendation systems
-    - **Feature Integration**: Better at combining diverse feature types
-    - **Scalability**: More efficient for large-scale recommendation tasks
-    - **Performance**: Higher accuracy for rating prediction tasks
-    - **Production Ready**: Optimized for real-time inference
-    
-    ### 💡 Best Use Cases:
-    - **Personalized Recommendations**: Based on user behavior and item characteristics
-    - **Rating Prediction**: Accurately predicts user preferences
-    - **Cold Start**: Handles new users and items through content features
-    - **Real-time Serving**: Fast inference for production systems
-    """)
-
-    with tab4:
-        st.header("📸 Book Gallery")
-        st.info("Browse book covers and discover new titles")
+    def get_user_recommendations(self, user_id: int, 
+                               candidate_books: List[str],
+                               k: int = 10,
+                               user_data: Optional[Dict] = None) -> List[Tuple[str, float]]:
+        """
+        Get top-k book recommendations for a user
         
-        # Gallery options
-        col1, col2 = st.columns([2, 1])
+        Args:
+            user_id: User ID
+            candidate_books: List of candidate book ISBNs
+            k: Number of recommendations
+            user_data: Additional user data
+            
+        Returns:
+            List of (book_isbn, prediction_score) tuples
+        """
+        if self.model is None or not self.torchrec_available:
+            print("❌ Model not loaded or torchrec not available")
+            return []
         
-        with col1:
-            gallery_mode = st.selectbox(
-                "Choose gallery mode",
-                ["Popular Books", "Recent Publications", "Random Selection", "Search Results"]
-            )
+        recommendations = []
         
-        with col2:
-            books_per_row = st.slider("Books per row", 2, 6, 4)
-            max_books = st.slider("Maximum books", 10, 50, 20)
+        print(f"Generating recommendations for user {user_id} from {len(candidate_books)} candidates...")
         
-        # Get books based on selected mode
-        if gallery_mode == "Popular Books":
-            # Get most rated books
-            book_popularity = ratings_df.groupby('ISBN').size().sort_values(ascending=False)
-            gallery_books = books_df[books_df['ISBN'].isin(book_popularity.head(max_books).index)]
-            
-        elif gallery_mode == "Recent Publications":
-            # Get recent books
-            books_df_temp = books_df.copy()
-            books_df_temp['Year-Of-Publication'] = pd.to_numeric(books_df_temp['Year-Of-Publication'], errors='coerce')
-            recent_books = books_df_temp.sort_values('Year-Of-Publication', ascending=False, na_position='last')
-            gallery_books = recent_books.head(max_books)
-            
-        elif gallery_mode == "Random Selection":
-            # Random books
-            gallery_books = books_df.sample(min(max_books, len(books_df)))
+        for book_isbn in candidate_books:
+            score = self.predict_rating(user_id, book_isbn, user_data)
+            recommendations.append((book_isbn, score))
+        
+        # Sort by score and return top-k
+        recommendations.sort(key=lambda x: x[1], reverse=True)
+        return recommendations[:k]
+    
+    def batch_recommend(self, user_ids: List[int], 
+                       candidate_books: List[str],
+                       k: int = 10) -> Dict[int, List[Tuple[str, float]]]:
+        """
+        Generate recommendations for multiple users
+        
+        Args:
+            user_ids: List of user IDs
+            candidate_books: List of candidate book ISBNs
+            k: Number of recommendations per user
             
-        else:  # Search Results
-            search_query = st.text_input("Search books for gallery", placeholder="Enter title, author, or publisher")
-            if search_query:
-                mask = (
-                    books_df['Book-Title'].str.contains(search_query, case=False, na=False) |
-                    books_df['Book-Author'].str.contains(search_query, case=False, na=False) |
-                    books_df['Publisher'].str.contains(search_query, case=False, na=False)
-                )
-                gallery_books = books_df[mask].head(max_books)
-            else:
-                gallery_books = books_df.head(max_books)
+        Returns:
+            Dictionary mapping user_id to recommendations
+        """
+        results = {}
+        
+        for user_id in user_ids:
+            results[user_id] = self.get_user_recommendations(user_id, candidate_books, k)
+        
+        return results
+    
+    def get_similar_books(self, target_book_isbn: str,
+                         candidate_books: List[str],
+                         sample_users: List[int],
+                         k: int = 10) -> List[Tuple[str, float]]:
+        """
+        Find books similar to target book by comparing user preferences
         
-        # Display gallery
-        if len(gallery_books) > 0:
-            st.markdown(f"**📚 Showing {len(gallery_books)} books**")
+        Args:
+            target_book_isbn: Target book ISBN
+            candidate_books: List of candidate book ISBNs
+            sample_users: Sample users to test similarity with
+            k: Number of similar books
             
-            # Create grid layout
-            books_list = gallery_books.to_dict('records')
+        Returns:
+            List of (book_isbn, similarity_score) tuples
+        """
+        target_scores = []
+        candidate_scores = {book: [] for book in candidate_books}
+        
+        # Get predictions for target book and candidates across sample users
+        for user_id in sample_users:
+            target_score = self.predict_rating(user_id, target_book_isbn)
+            target_scores.append(target_score)
             
-            # Display books in rows
-            for i in range(0, len(books_list), books_per_row):
-                cols = st.columns(books_per_row)
-                
-                for j, col in enumerate(cols):
-                    if i + j < len(books_list):
-                        book = books_list[i + j]
-                        
-                        with col:
-                            # Book cover
-                            image_url = book.get('Image-URL-M', '')
-                            
-                            if image_url and pd.notna(image_url) and str(image_url) != 'nan':
-                                try:
-                                    clean_url = str(image_url).strip()
-                                    if clean_url and 'http' in clean_url:
-                                        st.image(clean_url, width='stretch')
-                                    else:
-                                        st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width='stretch')
-                                except:
-                                    st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width='stretch')
-                            else:
-                                st.image("https://via.placeholder.com/150x200?text=📚&color=1f77b4&bg=f0f2f6", width='stretch')
-                            
-                            # Book info
-                            title = book['Book-Title']
-                            if len(title) > 40:
-                                title = title[:37] + "..."
-                            
-                            author = book['Book-Author']
-                            if len(author) > 25:
-                                author = author[:22] + "..."
-                            
-                            st.markdown(f"**{title}**")
-                            st.write(f"*{author}*")
-                            st.write(f"📅 {book.get('Year-Of-Publication', 'Unknown')}")
-                            
-                            # Book statistics
-                            book_stats = ratings_df[ratings_df['ISBN'] == book['ISBN']]
-                            if len(book_stats) > 0:
-                                avg_rating = book_stats['Book-Rating'].mean()
-                                num_ratings = len(book_stats)
-                                st.write(f"⭐ {avg_rating:.1f}/10 ({num_ratings} ratings)")
-                            else:
-                                st.write("⭐ No ratings")
-                            
-                            # DLRM prediction button
-                            if recommender and recommender.model:
-                                if st.button(f"🎯 DLRM Score", key=f"dlrm_{book['ISBN']}"):
-                                    with st.spinner("Calculating..."):
-                                        # Use first user as example
-                                        sample_user = users_df['User-ID'].iloc[0]
-                                        dlrm_score = recommender.predict_rating(sample_user, book['ISBN'])
-                                        st.success(f"DLRM Score: {dlrm_score:.3f}")
-        else:
-            st.info("No books found for the selected criteria")
+            for book_isbn in candidate_books:
+                if book_isbn != target_book_isbn:
+                    score = self.predict_rating(user_id, book_isbn)
+                    candidate_scores[book_isbn].append(score)
         
-        # Quick stats
-        st.markdown("---")
-        st.subheader("📊 Gallery Statistics")
+        # Calculate similarity based on correlation of user preferences
+        similarities = []
+        target_scores = np.array(target_scores)
         
-        col1, col2, col3, col4 = st.columns(4)
+        for book_isbn, scores in candidate_scores.items():
+            if len(scores) > 0:
+                scores_array = np.array(scores)
+                # Calculate correlation as similarity measure
+                correlation = np.corrcoef(target_scores, scores_array)[0, 1]
+                if not np.isnan(correlation):
+                    similarities.append((book_isbn, correlation))
         
-        with col1:
-            books_with_covers = sum(1 for _, book in gallery_books.iterrows()
-                                  if book.get('Image-URL-M') and pd.notna(book.get('Image-URL-M')))
-            st.metric("Books with Covers", f"{books_with_covers}/{len(gallery_books)}")
+        # Sort by similarity and return top-k
+        similarities.sort(key=lambda x: x[1], reverse=True)
+        return similarities[:k]
+
+
+def load_dlrm_recommender(model_source: str = "latest") -> DLRMBookRecommender:
+    """
+    Load DLRM recommender from various sources
+    
+    Args:
+        model_source: "latest" for latest MLflow run, "file" for local file, or specific run_id
         
-        with col2:
-            # Convert Year-Of-Publication to numeric, coercing errors to NaN
-            years = pd.to_numeric(gallery_books['Year-Of-Publication'], errors='coerce')
-            avg_year = years.mean()
-            st.metric("Average Publication Year", f"{avg_year:.0f}" if not pd.isna(avg_year) else "Unknown")
+    Returns:
+        DLRMBookRecommender instance
+    """
+    # Check if we're in CPU-only mode
+    cpu_only = os.environ.get('CPU_ONLY', 'false').lower() == 'true'
+    if cpu_only:
+        print("🔄 Loading recommender in CPU-only mode")
+    
+    # Create recommender instance
+    recommender = DLRMBookRecommender()
+    
+    # If torchrec is not available, return limited recommender
+    if not TORCHREC_AVAILABLE:
+        print("⚠️ torchrec not available, returning limited recommender")
+        return recommender
+    
+    if model_source == "latest":
+        # Try to get latest MLflow run
+        try:
+            experiment = mlflow.get_experiment_by_name('dlrm-book-recommendation-book_recommender')
+            if experiment:
+                runs = mlflow.search_runs(experiment_ids=[experiment.experiment_id],
+                                        order_by=["start_time desc"], max_results=1)
+                if len(runs) > 0:
+                    latest_run_id = runs.iloc[0].run_id
+                    recommender = DLRMBookRecommender(run_id=latest_run_id)
+                    return recommender
+        except Exception as e:
+            print(f"⚠️ Error loading from MLflow: {e}")
+    
+    elif model_source == "file":
+        # Try to load from local file
+        for filename in [
+'/home/mr-behdadi/PROJECT/ICE/notebooks/dlrm_book_model_final.pth',
+'/home/mr-behdadi/PROJECT/ICE/notebooks/dlrm_book_model_epoch_2.pth',
+'/home/mr-behdadi/PROJECT/ICE/notebooks/dlrm_book_model_epoch_0.pth',
+'/home/mr-behdadi/PROJECT/ICE/notebooks/dlrm_book_model_epoch_1.pth']:
+            if os.path.exists(filename):
+                try:
+                    recommender = DLRMBookRecommender(model_path=filename)
+                    return recommender
+                except Exception as e:
+                    print(f"⚠️ Error loading from {filename}: {e}")
+    
+    else:
+        # Treat as run_id
+        try:
+            recommender = DLRMBookRecommender(run_id=model_source)
+            return recommender
+        except Exception as e:
+            print(f"⚠️ Error loading from run_id {model_source}: {e}")
+    
+    print("⚠️ Could not load any trained model")
+    return recommender
+
+
+def demo_dlrm_recommendations():
+    """Demo function to show DLRM recommendations"""
+    
+    print("🚀 DLRM Book Recommendation Demo")
+    print("=" * 50)
+    
+    # Load book data for demo
+    books_df = pd.read_csv('Books.csv', encoding='latin-1', low_memory=False)
+    users_df = pd.read_csv('Users.csv', encoding='latin-1', low_memory=False)
+    ratings_df = pd.read_csv('Ratings.csv', encoding='latin-1', low_memory=False)
+    
+    books_df.columns = books_df.columns.str.replace('"', '')
+    users_df.columns = users_df.columns.str.replace('"', '')
+    ratings_df.columns = ratings_df.columns.str.replace('"', '')
+    
+    # Load recommender
+    recommender = load_dlrm_recommender("file")
+    
+    if recommender.model is None:
+        print("❌ No trained model found. Please run training first.")
+        return
+    
+    # Get sample user and books
+    sample_user_id = ratings_df['User-ID'].iloc[0]
+    sample_books = books_df['ISBN'].head(20).tolist()
+    
+    print(f"\n📚 Getting recommendations for User {sample_user_id}")
+    print(f"Testing with {len(sample_books)} candidate books...")
+    
+    # Get recommendations
+    recommendations = recommender.get_user_recommendations(
+        user_id=sample_user_id,
+        candidate_books=sample_books,
+        k=10
+    )
+    
+    print(f"\n🎯 Top 10 DLRM Recommendations:")
+    print("-" * 50)
+    
+    for i, (book_isbn, score) in enumerate(recommendations, 1):
+        # Get book info
+        book_info = books_df[books_df['ISBN'] == book_isbn]
+        if len(book_info) > 0:
+            book = book_info.iloc[0]
+            title = book['Book-Title']
+            author = book['Book-Author']
+            print(f"{i:2d}. {title} by {author}")
+            print(f"    ISBN: {book_isbn}, Score: {score:.4f}")
+        else:
+            print(f"{i:2d}. ISBN: {book_isbn}, Score: {score:.4f}")
+        print()
+    
+    # Show user's actual ratings for comparison
+    user_ratings = ratings_df[ratings_df['User-ID'] == sample_user_id]
+    if len(user_ratings) > 0:
+        print(f"\n📖 User {sample_user_id}'s Actual Reading History:")
+        print("-" * 50)
+        
+        for _, rating in user_ratings.head(5).iterrows():
+            book_info = books_df[books_df['ISBN'] == rating['ISBN']]
+            if len(book_info) > 0:
+                book = book_info.iloc[0]
+                print(f"• {book['Book-Title']} by {book['Book-Author']} - Rating: {rating['Book-Rating']}/10")
+    
+    # Test book similarity
+    if len(recommendations) > 0:
+        target_book = recommendations[0][0]
+        print(f"\n🔍 Finding books similar to: {target_book}")
         
-        with col3:
-            unique_authors = gallery_books['Book-Author'].nunique()
-            st.metric("Unique Authors", unique_authors)
+        similar_books = recommender.get_similar_books(
+            target_book_isbn=target_book,
+            candidate_books=sample_books,
+            sample_users=ratings_df['User-ID'].head(10).tolist(),
+            k=5
+        )
         
-        with col4:
-            unique_publishers = gallery_books['Publisher'].nunique()
-            st.metric("Unique Publishers", unique_publishers)
+        print(f"\n📚 Similar Books:")
+        print("-" * 30)
+        for i, (book_isbn, similarity) in enumerate(similar_books, 1):
+            book_info = books_df[books_df['ISBN'] == book_isbn]
+            if len(book_info) > 0:
+                book = book_info.iloc[0]
+                print(f"{i}. {book['Book-Title']} (similarity: {similarity:.3f})")
 
 if __name__ == "__main__":
-    main()
+    demo_dlrm_recommendations()