def add_args()

in src/fairseq/fairseq/models/wav2vec.py [0:0]

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

    def add_args(parser):
        """Add model-specific arguments to the parser."""
        parser.add_argument(
            "--prediction-steps",
            type=int,
            metavar="N",
            help="number of steps ahead to predict",
        )
        parser.add_argument(
            "--sample-distance",
            type=int,
            metavar="N",
            help="sample distance from target. does not work properly with cross-sampling",
        )
        parser.add_argument(
            "--cross-sample-negatives",
            type=int,
            metavar="N",
            help="num of cross sampled negatives",
        )
        parser.add_argument(
            "--num-negatives", type=int, metavar="N", help="number of negative examples"
        )
        parser.add_argument(
            "--conv-feature-layers",
            type=str,
            metavar="EXPR",
            help="convolutional feature extraction layers [(dim, kernel_size, stride), ...]",
        )
        parser.add_argument(
            "--conv-aggregator-layers",
            type=str,
            metavar="EXPR",
            help="convolutional feature extraction layers [(dim, kernel_size, stride), ...]",
        )
        parser.add_argument(
            "--dropout",
            type=float,
            metavar="D",
            help="dropout to apply within the model",
        )
        parser.add_argument(
            "--dropout-features",
            type=float,
            metavar="D",
            help="dropout to apply to the features",
        )
        parser.add_argument(
            "--dropout-agg",
            type=float,
            metavar="D",
            help="dropout to apply after aggregation step",
        )
        parser.add_argument(
            "--encoder", type=str, choices=["cnn"], help="type of encoder to use"
        )
        parser.add_argument(
            "--aggregator",
            type=str,
            choices=["cnn", "gru"],
            help="type of aggregator to use",
        )
        parser.add_argument(
            "--gru-dim", type=int, metavar="N", help="GRU dimensionality"
        )

        parser.add_argument(
            "--no-conv-bias",
            action="store_true",
            help="if set, does not learn bias for conv layers",
        )
        parser.add_argument(
            "--agg-zero-pad",
            action="store_true",
            help="if set, zero pads in aggregator instead of repl pad",
        )

        parser.add_argument(
            "--skip-connections-feat",
            action="store_true",
            help="if set, adds skip connections to the feature extractor",
        )
        parser.add_argument(
            "--skip-connections-agg",
            action="store_true",
            help="if set, adds skip connections to the aggregator",
        )
        parser.add_argument(
            "--residual-scale",
            type=float,
            metavar="D",
            help="scales residual by sqrt(value)",
        )

        parser.add_argument(
            "--log-compression",
            action="store_true",
            help="if set, adds a log compression to feature extractor",
        )

        parser.add_argument(
            "--balanced-classes",
            action="store_true",
            help="if set, loss is scaled to balance for number of negatives",
        )

        parser.add_argument(
            "--project-features",
            choices=["none", "same", "new"],
            help="if not none, features are projected using the (same or new) aggregator",
        )

        parser.add_argument(
            "--non-affine-group-norm",
            action="store_true",
            help="if set, group norm is not affine",
        )

        parser.add_argument(
            "--offset",
            help="if set, introduces an offset from target to predictions. "
            'if set to "auto", it is computed automatically from the receptive field',
        )

        parser.add_argument(
            "--activation",
            type=str,
            choices=["relu", "gelu"],
            help="which activation function to use",
        )

        parser.add_argument(
            "--vq-type",
            type=str,
            choices=["none", "gumbel", "kmeans"],
            help="which type of quantizer to use",
        )
        parser.add_argument(
            "--vq-vars",
            type=int,
            metavar="N",
            help="if set, project to this many vector quantized variables per group",
        )
        parser.add_argument(
            "--vq-groups",
            type=int,
            metavar="N",
            help="number of groups of latent variables",
        )
        parser.add_argument(
            "--vq-dim",
            type=int,
            metavar="N",
            help="uses this dimensionality for quantized vectors",
        )
        parser.add_argument(
            "--vq-depth",
            type=int,
            metavar="N",
            help="number of layers for vq weight projection",
        )
        parser.add_argument(
            "--combine-groups",
            action="store_true",
            help="if set, variables are shared among groups",
        )
        parser.add_argument(
            "--vq-temp",
            type=str,
            metavar="TEMP",
            help="temperature for latent variable sampling with gumbel softmax. should be a tuple of 3 values (start, end, decay)",
        )
        parser.add_argument(
            "--vq-gamma",
            type=float,
            metavar="D",
            help="gamma parameter for kmeans style vector quantization",
        )