classification.py

import pandas as pd
from transformers import AutoTokenizer, AutoModel
from sentence_transformers import SentenceTransformer, util
import numpy as np
import torch

def load_data(file_obj):
    # Assuming file_obj is a file-like object uploaded via Gradio, use `pd.read_excel` directly on it
    return pd.read_excel(file_obj)


def initialize_models():
    model_ST = SentenceTransformer("all-mpnet-base-v2")
    return model_ST


def generate_embeddings(df, model, Column):
    embeddings_list = []
    for index, row in df.iterrows():
        if type(row["Title"]) == str and type(row[Column]) == str:
            print(index)
            content = row["Title"] + "\n" + row[Column]
            embeddings = model.encode(content, convert_to_tensor=True)
            embeddings_list.append(embeddings)
        else:
            embeddings_list.append(np.nan)
    df['Embeddings'] = embeddings_list
    return df


def process_categories(categories, model):
    # Create a new DataFrame to store category information and embeddings
    df_cate = pd.DataFrame(categories)
    
    # Generate embeddings for each category description
    df_cate['Embeddings'] = df_cate.apply(lambda cat: model.encode(cat['description'], convert_to_tensor=True), axis=1)

    return df_cate



def match_categories(df, category_df):
    
    categories_list, experts_list, topic_list, scores_list = [], [], [], []
    for ebd_content in df['Embeddings']:
        if isinstance(ebd_content, torch.Tensor):
            cos_scores = util.cos_sim(ebd_content, torch.stack(list(category_df['Embeddings']), dim=0))[0]
            high_score_indices = [i for i, score in enumerate(cos_scores) if score > 0.45]

            # Append the corresponding categories, experts, and topics for each high-scoring index
            categories_list.append([category_df.loc[index, 'description'] for index in high_score_indices])
            experts_list.append([category_df.loc[index, 'experts'] for index in high_score_indices])
            topic_list.append([category_df.loc[index, 'topic'] for index in high_score_indices])
            scores_list.append([float(cos_scores[index]) for index in high_score_indices])
        else:
            categories_list.append(np.nan)
            experts_list.append(np.nan)
            topic_list.append(np.nan)
            scores_list.append('pas interessant')

    df["Description"] = categories_list
    df["Expert"] = experts_list
    df["Topic"] = topic_list
    df["Score"] = scores_list
    return df

def flatten_nested_lists(nested_list):
    """Flatten a list of potentially nested lists into a single list."""
    flattened_list = []
    for item in nested_list:
        if isinstance(item, list):
            flattened_list.extend(flatten_nested_lists(item))  # Recursively flatten the list
        else:
            flattened_list.append(item)
    return flattened_list

def save_data(df, filename):
    # Apply flattening and then join for the 'Expert' column
    df['Expert'] = df['Expert'].apply(lambda x: ', '.join(flatten_nested_lists(x)) if isinstance(x, list) else x)
    df['Description'] = df['Description'].apply(lambda x: ', '.join(x) if isinstance(x, list) else x)
    df['Topic'] = df['Topic'].apply(lambda x: ', '.join(x) if isinstance(x, list) else x)
    df['Score'] = df['Score'].apply(lambda x: ', '.join(map(str, x)) if isinstance(x, list) else x)


    new_filename = filename.replace(".", "_classified.")
    df.to_excel(new_filename, index=False)
    return new_filename

def classification(column, file_path, categories):
    # Load data
    df = load_data(file_path)

    # Initialize models
    model_ST = initialize_models()

    # Generate embeddings for df
    df = generate_embeddings(df, model_ST, column)


    category_df = process_categories(categories, model_ST)

    # Match categories
    df = match_categories(df, category_df)

    # Save data
    return save_data(df,file_path), df