python/preprocess_movies.py

import json
import os
import pandas as pd
from datasets import load_dataset, Dataset
from transformers import AutoTokenizer

# High-Level Utility Functions
def extract_unique_genres(df):
    """Extracts unique genres from the dataframe."""
    genre_list = []
    for genres in df['genre']:
        for genre in genres.split(','):
            genre_list.append(genre.strip())
    return sorted(list(set(genre_list)))

def encode_genres(genre_str):
    """Encodes the genres for multi-label classification."""
    genres = [g.strip() for g in genre_str.split(',')]
    encoding = [1 if genre in genres else 0 for genre in genre_to_int.keys()]
    return encoding

def tokenize_function(examples):
    """Tokenizes the text and returns only the necessary fields for the model."""
    encodings = tokenizer(examples["overview"], padding="max_length", truncation=True)
    encodings["labels"] = [encode_genres(genre_str) for genre_str in examples["genre"]]
    return encodings

# Load the dataset directly from HuggingFace DatasetHub
# Note: Adjust the token as per your HuggingFace access configuration
dataset = load_dataset('gatesfoundation/imdb-movies', data_files='data/imdb_movies.csv', split='train')

# Convert dataset to pandas dataframe
df = dataset.to_pandas()
df = df.dropna(subset=['overview', 'genre'])

# Extract unique genres
unique_genres = extract_unique_genres(df)

# Load or save genre_to_int mapping
mapping_file_path = 'genre_to_int.json'
if os.path.exists(mapping_file_path):
    with open(mapping_file_path, 'r') as f:
        genre_to_int = json.load(f)
else:
    genre_to_int = {genre: index for index, genre in enumerate(unique_genres)}
    with open(mapping_file_path, 'w') as f:
        json.dump(genre_to_int, f)

# Encode genres for multi-label classification
df['encoded_genres'] = df['genre'].apply(encode_genres)

# Convert the pandas dataframe back to a HuggingFace dataset
dataset = Dataset.from_pandas(df)

# Tokenize the dataset
tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
tokenized_datasets = dataset.map(tokenize_function, batched=True)

# Save the tokenized dataset in HuggingFace format
tokenized_datasets.save_to_disk('data/imdb_movies_preprocessed')