89 lines
3.7 KiB
Python
89 lines
3.7 KiB
Python
import click
|
|
from .model import TinyLanguageModel
|
|
from .helpers import read_mail_text, join
|
|
from .tokenization import tokenize_text, tokenize_words
|
|
|
|
|
|
@click.group()
|
|
def cli():
|
|
"""TinyLM - A simple n-gram language model."""
|
|
pass
|
|
|
|
|
|
@cli.command()
|
|
@click.option('-l', "--length", default=50, help="Number of tokens to generate.")
|
|
@click.option('-n', "--context-window-words", default=2, help="Number of words in the context window.")
|
|
@click.option('-f', "--filepath", type=click.Path(exists=True), multiple=True, help="Text file(s) to use as training corpus. Can be specified multiple times.")
|
|
@click.option('-g', "--gutenberg", multiple=True, help="NLTK Gutenberg corpus key(s). Can be specified multiple times.")
|
|
@click.option('-G', "--list-gutenberg", is_flag=True, help="List available Gutenberg corpus keys.")
|
|
@click.option('-m', "--mbox", type=click.Path(exists=True), help="Mbox file to use for training.")
|
|
@click.option('-p', "--prompt", help="Prompt to start generation.")
|
|
@click.option('-i', "--interact", is_flag=True, help="Drop into interactive shell after generating.")
|
|
@click.option('-t', "--tokenize", 'tokenize_opts', multiple=True, type=click.Choice(['lower', 'char', 'alpha']), help="Preprocessing option (can be specified multiple times). 'lower': lowercase all input text. 'char': use characters as tokens instead of words.")
|
|
@click.option('-v', "--verbose", is_flag=True, help="Display step-by-step generation as a table.")
|
|
def generate(length, context_window_words, filepath, gutenberg, list_gutenberg, mbox, prompt, interact, tokenize_opts, verbose):
|
|
"""Generate text using the language model."""
|
|
import nltk
|
|
|
|
# Handle --list-gutenberg: list available keys
|
|
if list_gutenberg:
|
|
nltk.download("gutenberg", quiet=True)
|
|
from nltk.corpus import gutenberg as gutenberg_corpus
|
|
click.echo("Available Gutenberg corpus keys:")
|
|
for key in gutenberg_corpus.fileids():
|
|
click.echo(f" {key}")
|
|
return
|
|
|
|
# Determine training corpus
|
|
corpus = []
|
|
|
|
if filepath:
|
|
for fp in filepath:
|
|
with open(fp, "r") as f:
|
|
corpus.extend(tokenize_text(f.read(), tokenize_opts))
|
|
if gutenberg:
|
|
nltk.download("gutenberg", quiet=True)
|
|
from nltk.corpus import gutenberg as gutenberg_corpus
|
|
for key in gutenberg:
|
|
corpus.extend(tokenize_words(gutenberg_corpus.words(key), tokenize_opts))
|
|
if mbox:
|
|
mail_text = read_mail_text(mbox)
|
|
corpus.extend(tokenize_text(mail_text, tokenize_opts))
|
|
|
|
if not corpus:
|
|
raise click.UsageError("No training data provided. Must specify at least one of --filepath, --gutenberg, or --mbox.")
|
|
|
|
# Train and generate
|
|
model = TinyLanguageModel(n=context_window_words)
|
|
model.train(corpus)
|
|
if prompt:
|
|
prompt_tokens = tokenize_text(prompt, tokenize_opts)
|
|
else:
|
|
prompt_tokens = None
|
|
join_fn = ''.join if 'char' in tokenize_opts else None
|
|
display_join = join_fn or join
|
|
|
|
if verbose:
|
|
from tabulate import tabulate
|
|
rows = []
|
|
import textwrap
|
|
def step_callback(pattern, options, chosen):
|
|
opts = textwrap.fill(', '.join(sorted(set(options))), width=60)
|
|
rows.append([display_join(list(pattern)), opts, chosen])
|
|
output = model.generate(length, prompt=prompt_tokens, join_fn=join_fn, step_callback=step_callback)
|
|
click.echo(tabulate(rows, headers=["Context", "Options", "Selected"], tablefmt="simple"))
|
|
click.echo()
|
|
|
|
else:
|
|
output = model.generate(length, prompt=prompt_tokens, join_fn=join_fn)
|
|
|
|
click.echo(output)
|
|
|
|
if interact:
|
|
import code
|
|
code.interact(local=locals(), banner="Entering interactive shell. 'model' and 'output' are available.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
cli()
|