def expand_generation_args()

in pytorch_translate/options.py [0:0]

• 121 lines of code
• 2 McCabe index (conditional complexity)

def expand_generation_args(group, train=False):
    """Expands the generation related arguments with pytorch_translate
    specific arguments"""
    group.add_argument(
        "--word-reward",
        type=float,
        default=0.0,
        help=(
            "Value to add to (log-prob) score for each token except EOS. "
            "Value < 0 encourages shorter translations, while > 0 "
            "(the usual case) encourages longer translations "
            "(similar to --length-penalty)."
        ),
    )
    group.add_argument(
        "--unk-reward",
        type=float,
        default=0.0,
        help=(
            "Value to add to (log-prob) score for UNK tokens. "
            "Value < 0 (the usual case) encourages fewer UNKs, while > 0 "
            "encourages more UNKs."
        ),
    )
    group.add_argument(
        "--length-penalty",
        type=float,
        default=0.0,
        help=(
            "When >0 scores are normalized according to length (divided by "
            "length^length_penalty). Effectively overrides word_reward when"
            "in use. NOTE: supersedes --lenpen."
        ),
    )
    group.add_argument(
        "--model-weights",
        default="",
        help=(
            "Interpolation weights for ensembles. Comma-separated list of "
            "floats with length equal to the number of models in the ensemble."
        ),
    )
    group.add_argument(
        "--report-oracle-bleu",
        type=utils.bool_flag,
        nargs="?",
        const=True,
        default=False,
        help=(
            "During evaluation, determine best among top-k outputs (where k "
            "is controlled by --nbest) for each sentence by smoothed "
            "sentence-level BLEU and report overall BLEU score for these "
            "sentences."
        ),
    )
    group.add_argument(
        "--output-hypos-binary-path",
        default=None,
        type=str,
        help=(
            "Optional filename to save output hypotheses (binary format "
            "and EOS-terminated, suitable for use as training targets)"
        ),
    )
    group.add_argument(
        "--max-examples-to-evaluate",
        default=-1,
        type=int,
        help=(
            "If >0 and smaller than size of evaluation data set, randomly "
            "sample this many examples to evaluate"
        ),
    )

    group.add_argument(
        "--max-examples-to-evaluate-seed",
        default=-1,
        type=int,
        help=(
            "If not -1, set seed for random sample as the given seed value, "
            "so we can replicate result on random sample."
        ),
    )

    group.add_argument(
        "--output-source-binary-path",
        default=None,
        type=str,
        help=(
            "Optional filename to save binarized source after evaluation "
            "(primary use case being that this source dataset will be after "
            "any filtering due to --max-examples-to-evaluate)"
        ),
    )
    group.add_argument(
        "--translation-info-export-path",
        default=None,
        type=str,
        help=("Optional path to save translation info output in pickled format"),
    )
    group.add_argument(
        "--diversity-sibling-gamma",
        type=float,
        default=0.0,
        help=("The diversity rate of sibling_rank for generating diverse beams"),
    )
    group.add_argument(
        "--hypotheses-export-path",
        default=None,
        type=str,
        help=("Optional path to save all generated hypotheses to external file"),
    )

    # These arguments are only used during training
    if train:
        group.add_argument(
            "--multi-model-restore-files",
            default=None,
            type=str,
            nargs="+",
            help=(
                "If --multi-encoder = --multi-decoder > 1, this option makes "
                "it possible to initialize individual model weights from "
                "existing checkpoints of separate training runs."
            ),
        )
    return group