Default values for val_steps_to_log and ar_steps_eval don't match

[sadamov]

As the title says, there are not enough ar_steps to log, by default, leading to an error on training start:

    parser.add_argument(
        "--ar_steps_eval",
        type=int,
        default=10,
        help="Number of steps to unroll prediction for during evaluation "
        "(default: 10)",
    )
    parser.add_argument(
        "--val_steps_to_log",
        nargs="+",
        type=int,
        default=[1, 2, 3, 5, 10, 15, 19],
        help="Steps to log val loss for (default: 1 2 3 5 10 15 19)",
    )

I suggest to change the val_steps_to_log to default=[1, 2, 3, 5, 7, 9] and its docstring.

[joeloskarsson]

This is the actual issue causing the error in #105. I suggest to close that and keep track of this here (better documented here).

[leifdenby]

I suggest to change the val_steps_to_log to default=[1, 2, 3, 5, 7, 9] and its docstring.

Rather than repeating the default variables in the help string I think we should use the ArgumentDefaultsHelpFormatter https://docs.python.org/3/library/argparse.html#argparse.ArgumentDefaultsHelpFormatter in argparse. Also I would make the two defaults be defined from a single list, something like:

DEFAULT_EVAL_STEPS_TO_LOG = [1, 2, 3, 5, 7, 9]
DEFAULT_NUM_EVAL_STEPS = max(DEFAULT_EVAL_STEPS_TO_LOG)

parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter, ...)
parser.add_argument(
        "--val_steps_to_log",
        nargs="+",
        type=int,
        default=[v for vin DEFAULT_VAL_STEPS if v <= DEFAULT_NUM_EVAL_STEPS,
        help="Steps to log val loss for",
    )

[joeloskarsson]

I've been thinking for ages that we should add ArgumentDefaultsHelpFormatter to all argparse uses, but keep forgetting to write up an issue about it 😅