app.py

import warnings
warnings.filterwarnings('ignore')

import streamlit as st
import pandas as pd
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from transformers import AutoTokenizer, AutoModel
import torch
from torch.utils.data import DataLoader, Dataset
from datasets import load_dataset  # For loading dataset
from datetime import datetime
from typing import List, Dict, Any
from functools import partial

# Configure GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Initialize session state
if 'history' not in st.session_state:
    st.session_state.history = []

if 'feedback' not in st.session_state:
    st.session_state.feedback = {}

# Define subset size and batch size for optimization
SUBSET_SIZE = 500  # Smaller subset for faster precomputation
BATCH_SIZE = 8  # Smaller batch size to reduce memory overhead

# Caching key resources: Model, Tokenizer, and Precomputed Embeddings
@st.cache_resource
def load_model_and_tokenizer():
    """
    Load the pre-trained model and tokenizer using Hugging Face Transformers.
    Cached to ensure it loads only once.
    """
    model_name = "Salesforce/codet5-small"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModel.from_pretrained(model_name).to(device)
    model.eval()
    return tokenizer, model

@st.cache_resource
def load_data():
    """
    Load and sample the dataset from Hugging Face.
    Ensures the 'text' column is created for embedding precomputation.
    """
    dataset = load_dataset("frankjosh/filtered_dataset")
    data = pd.DataFrame(dataset['train'])
    
    # Take a random subset of data
    data = data.sample(n=min(SUBSET_SIZE, len(data)), random_state=42).reset_index(drop=True)
    
    # Create a 'text' column by combining relevant fields
    data['text'] = data['docstring'].fillna('') + ' ' + data['summary'].fillna('')
    return data


@st.cache_resource
def precompute_embeddings(data: pd.DataFrame, _tokenizer, _model, batch_size=BATCH_SIZE):
    """
    Precompute embeddings for repository metadata to optimize query performance.
    The tokenizer and model are excluded from caching as they are unhashable.
    """
    class TextDataset(Dataset):
        def __init__(self, texts: List[str], tokenizer, max_length=512):
            self.texts = texts
            self.tokenizer = tokenizer
            self.max_length = max_length

        def __len__(self):
            return len(self.texts)

        def __getitem__(self, idx):
            return self.tokenizer(
                self.texts[idx],
                padding='max_length',
                truncation=True,
                max_length=self.max_length,
                return_tensors="pt"
            )

    def collate_fn(batch, pad_token_id):
        max_length = max(inputs['input_ids'].shape[1] for inputs in batch)
        input_ids, attention_mask = [], []
        for inputs in batch:
            input_ids.append(torch.nn.functional.pad(
                inputs['input_ids'].squeeze(),
                (0, max_length - inputs['input_ids'].shape[1]),
                value=pad_token_id
            ))
            attention_mask.append(torch.nn.functional.pad(
                inputs['attention_mask'].squeeze(),
                (0, max_length - inputs['attention_mask'].shape[1]),
                value=0
            ))
        return {
            'input_ids': torch.stack(input_ids),
            'attention_mask': torch.stack(attention_mask)
        }

    def generate_embeddings_batch(model, batch, device):
        with torch.no_grad():
            batch = {k: v.to(device) for k, v in batch.items()}
            outputs = model.encoder(**batch)
            return outputs.last_hidden_state.mean(dim=1).cpu().numpy()

    dataset = TextDataset(data['text'].tolist(), _tokenizer)
    dataloader = DataLoader(
        dataset, batch_size=batch_size, shuffle=False,
        collate_fn=partial(collate_fn, pad_token_id=_tokenizer.pad_token_id)
    )

    embeddings = []
    for batch in dataloader:
        batch_embeddings = generate_embeddings_batch(_model, batch, device)
        embeddings.extend(batch_embeddings)

    data['embedding'] = embeddings
    return data

@torch.no_grad()
def generate_query_embedding(model, tokenizer, query: str) -> np.ndarray:
    """
    Generate embedding for a user query using the pre-trained model.
    """
    inputs = tokenizer(
        query, return_tensors="pt", padding=True,
        truncation=True, max_length=512
    ).to(device)
    outputs = model.encoder(**inputs)
    return outputs.last_hidden_state.mean(dim=1).cpu().numpy()

def find_similar_repos(query_embedding: np.ndarray, data: pd.DataFrame, top_n=5) -> pd.DataFrame:
    """
    Compute cosine similarity and return the top N most similar repositories.
    """
    similarities = cosine_similarity([query_embedding], np.stack(data['embedding'].values))[0]
    data['similarity'] = similarities
    return data.nlargest(top_n, 'similarity')

def display_recommendations(recommendations: pd.DataFrame):
    """
    Display the recommended repositories in the Streamlit app interface.
    """
    st.markdown("### 🎯 Top Recommendations")
    for idx, row in recommendations.iterrows():
        st.markdown(f"### {idx + 1}. {row['repo']}")
        st.metric("Match Score", f"{row['similarity']:.2%}")
        st.markdown(f"[View Repository]({row['url']})")

# Main workflow
st.title("Repository Recommender System 🚀")
st.caption("Find repositories based on your project description.")

# Load resources
tokenizer, model = load_model_and_tokenizer()
data = load_data()
data = precompute_embeddings(data, tokenizer, model)

# User input
user_query = st.text_area(
    "Describe your project:", height=150,
    placeholder="Example: A machine learning project for customer churn prediction..."
)

if st.button("🔍 Search Repositories"):
    if user_query.strip():
        with st.spinner("Finding relevant repositories..."):
            query_embedding = generate_query_embedding(model, tokenizer, user_query)
            recommendations = find_similar_repos(query_embedding, data)
            display_recommendations(recommendations)
    else:
        st.error("Please provide a project description.")