cli.py

#!/usr/bin/env python3
import warnings
# The warnings come from upstream (Haiku) and would only clutter stderr.
warnings.filterwarnings('ignore')
from typing import Optional
from pathlib import Path
import os
import click
import jax

from kigo.utils import File, get_logger, get_rngs
from kigo import nn
from kigo import diffusion
from kigo import viz
from kigo import persistence
from kigo import training
from kigo import data


logger = get_logger('kigo.cli')


@click.group('Kigo')
def cli() -> None:
    '''A diffusion model written with DM-Haiku!'''
    pass


@cli.command('init')
@click.argument('workdir', type=Path)
@click.argument('config', type=File)
@click.option('--seed', type=int, default=None)
def init(workdir: Path, config: File, seed: Optional[int]) -> None:
    '''Initializes a new model in the given working directory with the given
    configuration file.'''
    rngs = get_rngs(seed)
    persistence.init_workdir(workdir, config, rngs)


@cli.command('train')
@click.argument('checkpoint', type=persistence.get_checkpoint)
@click.option('--debug', is_flag=True)
@click.option('--seed', type=int, default=None)
def train(checkpoint: Path, debug: bool, seed: Optional[int]) -> None:
    '''Trains the model in the given directory. The latter can either be a
    working directory as specified in the `init` command or a checkpoint within
    a working directory. The training continues until interrupted with
    Ctrl+c.'''
    jax.config.update('jax_disable_jit', debug)  # type: ignore
    rngs = get_rngs(seed)
    cfg = persistence.load_cfg(checkpoint)
    params = persistence.load_params(checkpoint)
    ema = persistence.load_ema(checkpoint)
    opt_state = persistence.load_opt_state(checkpoint)
    dataset = data.Dataset.from_cfg(cfg)
    workdir = persistence.get_workdir(checkpoint)
    ctx = persistence.load_ctx(checkpoint)
    packs = training.train(params, ema, opt_state, rngs, dataset, workdir,
                           cfg, ctx)
    packs = training.autosave(packs)
    if cfg.tr.wandb:
        packs = training.wandb_log(packs)
    packs = training.log(packs)
    for pack in packs:
        del pack


@cli.command('syn')
@click.argument('checkpoint', type=persistence.get_checkpoint)
@click.option('--steps', type=int, default=64)
@click.option('--no-ema', is_flag=True)
@click.option('--eta', type=float, default=0.)
@click.option('--clip-percentile', '-c', type=float, default=0.995)
@click.option('--out', '-o', type=Path, default=None)
@click.option('--number', '-n', type=int, default=1)
@click.option('--debug', is_flag=True)
@click.option('--seed', type=int, default=None)
def syn(checkpoint: Path, steps: int, no_ema: bool, eta: float,
        clip_percentile: float, out: Optional[Path], number: int,
        seed: Optional[int], debug: bool) -> None:
    '''Synthesizes a new image from the model using DDIM sampling.'''
    jax.config.update('jax_disable_jit', debug)  # type: ignore
    rngs = get_rngs(seed)
    cfg = persistence.load_cfg(checkpoint)
    params = (persistence.load_params(checkpoint)
              if no_ema else
              persistence.load_ema(checkpoint))
    _, forward_fn = nn.get_params_and_forward_fn(cfg, rngs, params)
    xT = jax.random.normal(next(rngs), shape=(number, *cfg.img.shape))
    x0 = diffusion.sample_p(xT, forward_fn, steps, next(rngs), eta,
                            clip_percentile)
    viz.show(x0, out=out)


@cli.command('slurm')
@click.argument('cp', type=persistence.get_checkpoint)
@click.option('--time', type=str, required=True)
@click.option('--email', type=str, required=True)
@click.option('--cpus', type=int, default=10)
@click.option('--memory', type=int, default=20)
@click.option('--gpus', type=int, default=1)
@click.option('--logfile', type=Path, default=None)
@click.option('--out', '-o', type=Path, default=None)
def slurm_cli(cp: Path,
              time: str,
              email: str,
              cpus: int,
              memory: int,
              gpus: int,
              out: Optional[Path],
              logfile: Optional[Path],
              ) -> None:
    '''Generates a Slurm script to train from a given checkpoint.'''
    with open('slurm.template.sh') as fh:
        template = fh.read()
    logfile = logfile or persistence.get_workdir(cp) / 'logs.log'
    out = out or persistence.get_workdir(cp) / 'slurm.sh'
    logfile.parent.mkdir(parents=True, exist_ok=True)
    ctx = dict(kigo_dir=os.getcwd(), checkpoint=str(cp), time=time,
               email=email, logfile=str(logfile), cpus=cpus, memory=memory,
               gpus=gpus)
    src = template % ctx
    with open(out, 'w') as fh:
        fh.write(src.strip())
    logger.info(f'Saved Slurm script to {out}')


if __name__ == '__main__':
    cli()