generate_data.py

import re
import json


def clean_text_file(input_file, output_file):
    """
    Cleans the text file by:
    - Removing lines that start with 'BOOK' or contain only numbers,
    - Substituting two leading spaces with a single tab,
    - Normalizing all leading whitespace to no more than two tabs of indentation,
    - Splitting the text into chunks separated by at least 3 newlines.
    """
    with open(input_file, "r", encoding="utf-8") as infile:
        lines = infile.readlines()

    cleaned_lines = []

    for line in lines:
        stripped_line = line.strip()
        if stripped_line.startswith("BOOK") or stripped_line.isdigit():
            continue

        line = re.sub(r"^ ", "\t", line)
        line = re.sub(r"^[\t ]+", normalize_whitespace, line)

        # Remove a single leading tab if present
        if line.startswith("\t"):
            line = line[1:]
        cleaned_lines.append(line)

    cleaned_text = "".join(cleaned_lines)
    chunks = [chunk.strip() for chunk in cleaned_text.split("\n\n\n") if chunk.strip()]

    with open(output_file, "w", encoding="utf-8") as outfile:
        outfile.writelines(cleaned_lines)

    return chunks


def normalize_whitespace(match):
    whitespace = match.group(0).replace("\t", "  ")
    tab_count = len(whitespace) // 2
    return "\t" * min(tab_count, 2)


def chunk_poem_by_lines(lines, max_chunk_lines, overlap_lines):
    """
    Split a poem into chunks of manageable size, preserving lines and overlap for continuity.
    """
    chunks = []
    for i in range(0, len(lines), max_chunk_lines - overlap_lines):
        chunk = lines[i : i + max_chunk_lines]
        chunks.append(chunk)
    return chunks


def generate_training_pairs(
    poems, max_context_lines=10, max_chunk_lines=20, overlap_lines=10
):
    """
    Generate input-output training pairs for poetry generation, using line-based chunking.
    max_context_lines: The maximum number of lines to consider as context for the next line.
    max_chunk_lines: The maximum number of lines to consider in a single chunk.
    overlap_lines: The number of lines to overlap between chunks for continuity.


    """
    training_data = []

    for poem in poems:
        lines = poem.splitlines(keepends=True)

        # Chunk the poem into manageable pieces
        chunks = chunk_poem_by_lines(
            lines, max_chunk_lines=max_chunk_lines, overlap_lines=overlap_lines
        )

        for chunk in chunks:
            for i in range(1, len(chunk)):
                input_lines = "".join(chunk[max(0, i - max_context_lines) : i])
                output_line = chunk[i]  # The next line to predict
                training_data.append({"input": input_lines, "output": output_line})

    return training_data


# Example usage
poems = clean_text_file("leaves-of-grass-original.txt", "leaves-of-grass-cleaned.txt")

# 400 poems
print(f"Number of poems: {len(poems)}")

# TODO: Implement generate_training_pairs
# poems = poems[25:26]  # For testing purposes. Long poem.


# Compact Poetry Model
max_context_lines = 5
max_chunk_lines = 10
overlap_lines = 2

"""
# Narrative Poetry Model
max_context_lines = 10
max_chunk_lines = 20
overlap_lines = 5
"""

"""
# Epic or Free Verse Poetry 
max_context_lines = 20
max_chunk_lines = 50
overlap_lines = 10
"""

training_data = generate_training_pairs(
    poems,
    max_context_lines=max_context_lines,
    max_chunk_lines=max_chunk_lines,
    overlap_lines=overlap_lines,
)

print(f"Number of training pairs: {len(training_data)}")

# write the training data to a file
with open("train.json", "w") as outfile:
    json.dump(training_data, outfile, indent=2, ensure_ascii=False)