def expand_optimization_args()

in pytorch_translate/options.py [0:0]

• 81 lines of code
• 1 McCabe index (conditional complexity)

def expand_optimization_args(group):
    """Expands the optimization related arguments with pytorch_translate
    specific arguments"""
    group.add_argument(
        "--local-num-gpus",
        default=torch.cuda.device_count(),
        type=int,
        metavar="N",
        help=(
            "The number of local GPUs to use for training on this machine. "
            "Defaults to using all visible GPUs. This should be "
            "<= --distributed-world-size."
        ),
    )
    group.add_argument(
        "--stop-time-hr",
        default=-1.0,
        type=float,
        metavar="N",
        help="Stops training after N hours have elapsed. Use decimal values "
        "for sub-hourly granularity. A value of < 0 disables this.",
    )
    group.add_argument(
        "--stop-no-best-validate-loss",
        default=-1,
        type=int,
        metavar="N",
        help="Stops training after N validations have been run without "
        "achieving a better loss than before. Note that this is affected by "
        "--save-interval-updates in how frequently we run validation in the "
        "first place. A value of < 0 disables this.",
    )
    group.add_argument(
        "--stop-no-best-bleu-eval",
        default=-1,
        type=int,
        metavar="N",
        help="Stops training after N evals have been run without "
        "achieving a better BLEU score than before. Note that this is affected "
        "by --save-interval-updates in how frequently we run BLEU eval "
        "in the first place. A value of < 0 disables this.",
    )
    group.add_argument(
        "--shrink-lr-no-best-tune-loss",
        default=5,
        type=int,
        metavar="N",
        help="Decay learning rate after N evals have been run without "
        "achieving a lower tune loss than before. This is to achieve "
        "decay lr within an epoch, independent of lr_scheduler. "
        "Note that this is affected by --save-interval-updates in "
        "how frequently we run BLEU eval in the first place. "
        "A value of < 0 disables this.",
    )
    group.add_argument(
        "--pruning-percentile",
        type=int,
        default=0,
        help="Proportion of weights to prune. A value <=0 disables pruning."
        " By default, prunes weights uniformly and ignores bias terms.",
    )
    group.add_argument(
        "--parameters-to-prune",
        default="all",
        help="Names of layers to prune. Layers are pruned if the argument is "
        "a substring of the layer name. Options are 'all', 'embed', 'lstm'. ",
    )
    group.add_argument(
        "--loss-beam",
        type=int,
        default=0,
        help="Beam size to use for 'sequence_nll' loss and 'sequence_risk' "
        "loss. If zero, use --beam.",
    )
    group.add_argument(
        "--disable-eval-bleu",
        nargs="?",
        const=True,
        default=False,
        type=utils.bool_flag,
        help=("disable bleu score evaluation on tune dataset"),
    )

    return group