def add(a, *args, default=None, func=add_argument, **kwargs):
arg = func(a,
*args,
default=default if arg_default is None else arg_default,
**kwargs)
try:
RESTORED_ARGS.add(arg.dest)
except AttributeError:
RESTORED_ARGS.update(get_group_arg_names(arg))
def add(a, *args, default=None, func=add_argument, **kwargs):
arg = func(a, *args, default=default if arg_default is None else arg_default, **kwargs)
try:
RESTORED_ARGS.add(arg.dest)
except AttributeError:
RESTORED_ARGS.update(get_group_arg_names(arg))
def add_param_arguments(ap=None, arg_default=None): # arguments with possible format-specific parameter values
def add_argument(a, *args, **kwargs):
return a.add_argument(*args, **kwargs)
def add(a, *args, default=None, func=add_argument, **kwargs):
arg = func(a, *args, default=default if arg_default is None else arg_default, **kwargs)
try:
RESTORED_ARGS.add(arg.dest)
except AttributeError:
RESTORED_ARGS.update(get_group_arg_names(arg))
def add_boolean(a, *args, **kwargs):
add(a, *args, func=add_boolean_option, **kwargs)
if not ap:
ap = ArgParser()
group = ap.add_argument_group(title="Node labels")
add(group, "--max-node-labels", type=int, default=0, help="max number of node labels to allow")
add(group, "--max-node-categories", type=int, default=0, help="max node categories to allow")
add(group, "--min-node-label-count", type=int, default=2, help="min number of occurrences for a label")
add_boolean(group, "use-gold-node-labels", "gold node labels when parsing")
add_boolean(group, "wikification", "use Spotlight to wikify any named node")
add_boolean(group, "node-labels", "prediction of node labels, if supported by format", default=True)
group = ap.add_argument_group(title="Structural constraints")
add_boolean(group, "linkage", "linkage nodes and edges")
add_boolean(group, "implicit", "implicit nodes and edges")
add_boolean(group, "remote", "remote edges", default=True)
add_boolean(group, "constraints", "scheme-specific rules", default=True)
add_boolean(group, "require-connected", "constraint that output graph must be connected")
add(group, "--orphan-label", default="orphan", help="edge label to use for nodes without parents")
add(group, "--max-action-ratio", type=float, default=100, help="max action/terminal ratio")
add(group, "--max-node-ratio", type=float, default=10, help="max node/terminal ratio")
add(group, "--max-height", type=int, default=20, help="max graph height")
group = ap.add_mutually_exclusive_group()
add(group, "--swap", choices=(REGULAR, COMPOUND), default=REGULAR, help="swap transitions")
add(group, "--no-swap", action="store_false", dest="swap", help="exclude swap transitions")
add(ap, "--max-swap", type=int, default=15, help="if compound swap enabled, maximum swap size")
group = ap.add_argument_group(title="General classifier training parameters")
add(group, "--learning-rate", type=float, help="rate for model weight updates (default: by trainer/1)")
add(group, "--learning-rate-decay", type=float, default=0, help="learning rate decay per iteration")
add(group, "--swap-importance", type=float, default=1, help="learning rate factor for Swap")
add(group, "--max-training-per-format", type=int, help="max number of training passages per format per iteration")
add_boolean(group, "missing-node-features", "allow node features to be missing if not available", default=True)
add(group, "--omit-features", help="string of feature properties to omit, out of " + FEATURE_PROPERTIES)
add_boolean(group, "curriculum", "sort training passages by action prediction accuracy in previous epoch")
group = ap.add_argument_group(title="Perceptron parameters")
add(group, "--min-update", type=int, default=5, help="minimum #updates for using a feature")
SPARSE_ARG_NAMES.update(get_group_arg_names(group))
group = ap.add_argument_group(title="Neural network parameters")
add(group, "--word-dim-external", type=int, default=300, help="dimension for external word embeddings")
add(group, "--word-vectors", help="file to load external word embeddings from (default: GloVe)")
add(group, "--vocab", help="file mapping integer ID to word form (to avoid loading spaCy), or '-' to use word form")
add_boolean(group, "update-word-vectors", "external word vectors in training parameters", default=True)
add(group, "--word-dim", type=int, default=0, help="dimension for learned word embeddings")
add(group, "--lemma-dim", type=int, default=200, help="dimension for lemma embeddings")
add(group, "--tag-dim", type=int, default=20, help="dimension for fine POS tag embeddings")
add(group, "--pos-dim", type=int, default=20, help="dimension for coarse/universal POS tag embeddings")
add(group, "--dep-dim", type=int, default=10, help="dimension for dependency relation embeddings")
add(group, "--edge-label-dim", type=int, default=20, help="dimension for edge label embeddings")
add(group, "--node-label-dim", type=int, default=0, help="dimension for node label embeddings")
add(group, "--node-category-dim", type=int, default=0, help="dimension for node category embeddings")
add(group, "--punct-dim", type=int, default=1, help="dimension for separator punctuation embeddings")
add(group, "--action-dim", type=int, default=3, help="dimension for input action type embeddings")
add(group, "--ner-dim", type=int, default=3, help="dimension for input entity type embeddings")
add(group, "--shape-dim", type=int, default=3, help="dimension for word shape embeddings")
add(group, "--prefix-dim", type=int, default=2, help="dimension for word prefix embeddings")
add(group, "--suffix-dim", type=int, default=3, help="dimension for word suffix embeddings")
add(group, "--output-dim", type=int, default=50, help="dimension for output action embeddings")
add(group, "--layer-dim", type=int, default=50, help="dimension for hidden layers")
add(group, "--layers", type=int, default=2, help="number of hidden layers")
add(group, "--lstm-layer-dim", type=int, default=500, help="dimension for LSTM hidden layers")
add(group, "--lstm-layers", type=int, default=0, help="number of LSTM hidden layers")
add(group, "--embedding-layer-dim", type=int, default=500, help="dimension for layers before LSTM")
add(group, "--embedding-layers", type=int, default=1, help="number of layers before LSTM")
add(group, "--activation", choices=ACTIVATIONS, default=DEFAULT_ACTIVATION, help="activation function")
add(group, "--init", choices=INITIALIZERS, default=DEFAULT_INITIALIZER, help="weight initialization")
add(group, "--minibatch-size", type=int, default=200, help="mini-batch size for optimization")
add(group, "--optimizer", choices=TRAINERS, default=DEFAULT_TRAINER, help="algorithm for optimization")
add(group, "--loss", choices=LOSSES, default=DEFAULT_LOSS, help="loss function for training")
add(group, "--max-words-external", type=int, default=250000, help="max external word vectors to use")
add(group, "--max-words", type=int, default=10000, help="max number of words to keep embeddings for")
add(group, "--max-lemmas", type=int, default=3000, help="max number of lemmas to keep embeddings for")
add(group, "--max-tags", type=int, default=100, help="max number of fine POS tags to keep embeddings for")
add(group, "--max-pos", type=int, default=100, help="max number of coarse POS tags to keep embeddings for")
add(group, "--max-deps", type=int, default=100, help="max number of dep labels to keep embeddings for")
add(group, "--max-edge-labels", type=int, default=15, help="max number of edge labels for embeddings")
add(group, "--max-puncts", type=int, default=5, help="max number of punctuations for embeddings")
add(group, "--max-action-types", type=int, default=10, help="max number of action types for embeddings")
add(group, "--max-action-labels", type=int, default=100, help="max number of action labels to allow")
add(group, "--max-ner-types", type=int, default=18, help="max number of entity types to allow")
add(group, "--max-shapes", type=int, default=30, help="max number of word shapes to allow")
add(group, "--max-prefixes", type=int, default=30, help="max number of 1-character word prefixes to allow")
add(group, "--max-suffixes", type=int, default=500, help="max number of 3-character word suffixes to allow")
add(group, "--word-dropout", type=float, default=0.2, help="word dropout parameter")
add(group, "--word-dropout-external", type=float, default=0, help="word dropout for word vectors")
add(group, "--lemma-dropout", type=float, default=0.2, help="lemma dropout parameter")
add(group, "--tag-dropout", type=float, default=0.2, help="fine POS tag dropout parameter")
add(group, "--pos-dropout", type=float, default=0.2, help="coarse POS tag dropout parameter")
add(group, "--dep-dropout", type=float, default=0.5, help="dependency label dropout parameter")
add(group, "--node-label-dropout", type=float, default=0.2, help="node label dropout parameter")
add(group, "--node-dropout", type=float, default=0.1, help="probability to drop features for a whole node")
add(group, "--dropout", type=float, default=0.4, help="dropout parameter between layers")
add(group, "--max-length", type=int, default=120, help="maximum length of input sentence")
add(group, "--rnn", choices=["None"] + list(RNNS), default=DEFAULT_RNN, help="type of recurrent neural network")
add(group, "--gated", type=int, nargs="?", default=2, help="gated input to BiRNN and MLP")
NN_ARG_NAMES.update(get_group_arg_names(group))
return ap
def __init__(self, *args):
self.arg_parser = ap = ArgParser(description="Transition-based parser for UCCA.",
formatter_class=ArgumentDefaultsHelpFormatter)
ap.add_argument("passages", nargs="*", help="passage files/directories to test on/parse")
ap.add_argument("--version", action="version", version="")
ap.add_argument("-C", "--config", is_config_file=True, help="configuration file to get arguments from")
ap.add_argument("-m", "--models", nargs="+", help="model file basename(s) to load/save, ensemble if >1 "
"(default: <format>_<model_type>")
ap.add_argument("-c", "--classifier", choices=CLASSIFIERS, default=BIRNN, help="model type")
ap.add_argument("-B", "--beam", type=int, choices=(1,), default=1, help="beam size for beam search")
add_boolean_option(ap, "evaluate", "evaluation of parsed passages", short="e")
add_verbose_arg(ap, help="detailed parse output")
constructions.add_argument(ap)
add_boolean_option(ap, "sentences", "split to sentences")
add_boolean_option(ap, "paragraphs", "split to paragraphs")
ap.add_argument("--timeout", type=float, help="max number of seconds to wait for a single passage")
group = ap.add_argument_group(title="Training parameters")
group.add_argument("-t", "--train", nargs="+", default=(), help="passage files/directories to train on")
group.add_argument("-d", "--dev", nargs="+", default=(), help="passage files/directories to tune on")
group.add_argument("-I", "--iterations", nargs="+", type=Iterations,
default=(Iterations(50), Iterations("100 --optimizer=" + EXTRA_TRAINER)),
help="number of training iterations along with optional hyperparameters per part")
group.add_argument("--folds", type=int, choices=(3, 5, 10), help="#folds for cross validation")
group.add_argument("--seed", type=int, default=1, help="random number generator seed")
add_boolean_option(group, "early-update", "early update procedure (finish example on first error)")
group.add_argument("--save-every", type=int, help="every this many passages, evaluate on dev and save model")
add_boolean_option(group, "eval-test", "evaluate on test whenever evaluating on dev, but keep results hidden")
add_boolean_option(group, "ignore-case", "pre-convert all input files to lower-case in training and test")
group = ap.add_argument_group(title="Output files")
group.add_argument("-o", "--outdir", default=".", help="output directory for parsed files")
group.add_argument("-p", "--prefix", default="", help="output filename prefix")
add_boolean_option(group, "write", "writing parsed output to files", default=True, short_no="W")
group.add_argument("-j", "--join", help="if output format is textual, write all to one file with this basename")
group.add_argument("-l", "--log", help="output log file (default: model filename + .log)")
group.add_argument("--devscores", help="output CSV file for dev scores (default: model filename + .dev.csv)")
group.add_argument("--testscores", help="output CSV file for test scores (default: model filename + .test.csv)")
group.add_argument("--action-stats", help="output CSV file for action statistics")
add_boolean_option(group, "normalize", "apply normalizations to output in case format is UCCA", default=False)
ap.add_argument("-f", "--formats", nargs="+", choices=FILE_FORMATS, default=(),
help="input formats for creating all parameters before training starts "
"(otherwise created dynamically based on filename suffix), "
"and output formats for written files (each will be written; default: UCCA XML)")
ap.add_argument("-u", "--unlabeled", nargs="*", choices=FORMATS, help="to ignore labels in")
ap.add_argument("--lang", default="en", help="two-letter language code to use as the default language")
add_boolean_option(ap, "multilingual", "separate model parameters per language (passage.attrib['lang'])")
group = ap.add_argument_group(title="Sanity checks")
add_boolean_option(group, "check-loops", "check for parser state loop")
add_boolean_option(group, "verify", "check for oracle reproducing original passage")
add_boolean_option(group, "validate-oracle", "require oracle output to respect constraints", default=True)
add_param_arguments(ap)
group = ap.add_argument_group(title="DyNet parameters")
group.add_argument("--dynet-mem", help="memory for dynet")
group.add_argument("--dynet-weight-decay", type=float, default=1e-5, help="weight decay for parameters")
add_boolean_option(group, "dynet-apply-weight-decay-on-load", "workaround for clab/dynet#1206", default=False)
add_boolean_option(group, "dynet-gpu", "GPU for training")
group.add_argument("--dynet-gpus", type=int, default=1, help="how many GPUs you want to use")
add_boolean_option(group, "dynet-autobatch", "auto-batching of training examples")
DYNET_ARG_NAMES.update(get_group_arg_names(group))
ap.add_argument("-H", "--hyperparams", type=HyperparamsInitializer.action, nargs="*",
help="shared hyperparameters or hyperparameters for specific formats, "
'e.g., "shared --lstm-layer-dim=100 --lstm-layers=1" "ucca --word-dim=300"',
default=[HyperparamsInitializer.action("shared --lstm-layers 2")])
ap.add_argument("--copy-shared", nargs="*", choices=FORMATS, help="formats whose parameters shall be "
"copied from loaded shared parameters")
self.args = FallbackNamespace(ap.parse_args(args if args else None))
if self.args.config:
print("Loading configuration from '%s'." % self.args.config)
if self.args.passages and self.args.write:
os.makedirs(self.args.outdir, exist_ok=True)
if self.args.models:
if not self.args.log:
self.args.log = self.args.models[0] + ".log"
if self.args.dev and not self.args.devscores:
self.args.devscores = self.args.models[0] + ".dev.csv"
if self.args.passages and not self.args.testscores:
self.args.testscores = self.args.models[0] + ".test.csv"
elif not self.args.log:
self.args.log = "parse.log"
self.sub_configs = [] # Copies to be stored in Models so that they do not interfere with each other
self._logger = self.format = self.hyperparams = self.iteration_hyperparams = None
self._vocab = {}
self.original_values = {}
self.random = np.random
self.update()
def __init__(self, *args):
self.arg_parser = ap = ArgParser(
description="Transition-based parser for UCCA.",
formatter_class=ArgumentDefaultsHelpFormatter)
add_boolean_option(ap,
"use-bert",
default=False,
description="whether to use bert embeddings")
ap.add_argument("--bert-model",
choices=[
"bert-base-uncased", "bert-large-uncased",
"bert-base-cased", "bert-large-cased",
"bert-base-multilingual-cased"
],
default="bert-base-multilingual-cased")
ap.add_argument("--bert-layers",
type=int,
nargs='+',
default=[-1, -2, -3, -4])
ap.add_argument("--bert-layers-pooling",
choices=["weighted", "sum", "concat"],
default="weighted")
ap.add_argument("--bert-token-align-by",
choices=["first", "sum", "mean"],
default="sum")
ap.add_argument("--bert-multilingual", choices=[0], type=int)
add_boolean_option(
ap,
"bert-use-default-word-embeddings",
default=False,
description="whether to use default word embeddings")
ap.add_argument("--bert-dropout",
type=float,
default=0,
choices=np.linspace(0, 0.9, num=10))
ap.add_argument("passages",
nargs="*",
help="passage files/directories to test on/parse")
ap.add_argument("--version", action="version", version="")
ap.add_argument("-C",
"--config",
is_config_file=True,
help="configuration file to get arguments from")
ap.add_argument(
"-m",
"--models",
nargs="+",
help="model file basename(s) to load/save, ensemble if >1 "
"(default: <format>_<model_type>")
ap.add_argument("-c",
"--classifier",
choices=CLASSIFIERS,
default=BIRNN,
help="model type")
ap.add_argument("-B",
"--beam",
type=int,
choices=(1, ),
default=1,
help="beam size for beam search")
add_boolean_option(ap,
"evaluate",
"evaluation of parsed passages",
short="e")
add_verbose_arg(ap, help="detailed parse output")
constructions.add_argument(ap)
add_boolean_option(ap, "sentences", "split to sentences")
add_boolean_option(ap, "paragraphs", "split to paragraphs")
ap.add_argument(
"--timeout",
type=float,
help="max number of seconds to wait for a single passage")
group = ap.add_argument_group(title="Training parameters")
group.add_argument("-t",
"--train",
nargs="+",
default=(),
help="passage files/directories to train on")
group.add_argument("-d",
"--dev",
nargs="+",
default=(),
help="passage files/directories to tune on")
group.add_argument(
"-I",
"--iterations",
nargs="+",
type=Iterations,
default=(Iterations(50),
Iterations("100 --optimizer=" + EXTRA_TRAINER)),
help=
"number of training iterations along with optional hyperparameters per part"
)
group.add_argument("--folds",
type=int,
choices=(3, 5, 10),
help="#folds for cross validation")
group.add_argument("--seed",
type=int,
default=1,
help="random number generator seed")
add_boolean_option(
group, "early-update",
"early update procedure (finish example on first error)")
group.add_argument(
"--save-every",
type=int,
help="every this many passages, evaluate on dev and save model")
add_boolean_option(
group, "eval-test",
"evaluate on test whenever evaluating on dev, but keep results hidden"
)
add_boolean_option(
group, "ignore-case",
"pre-convert all input files to lower-case in training and test")
group = ap.add_argument_group(title="Output files")
group.add_argument("-o",
"--outdir",
default=".",
help="output directory for parsed files")
group.add_argument("-p",
"--prefix",
default="",
help="output filename prefix")
add_boolean_option(group,
"write",
"writing parsed output to files",
default=True,
short_no="W")
group.add_argument(
"-j",
"--join",
help=
"if output format is textual, write all to one file with this basename"
)
group.add_argument(
"-l",
"--log",
help="output log file (default: model filename + .log)")
group.add_argument(
"--devscores",
help=
"output CSV file for dev scores (default: model filename + .dev.csv)"
)
group.add_argument(
"--testscores",
help=
"output CSV file for test scores (default: model filename + .test.csv)"
)
group.add_argument("--action-stats",
help="output CSV file for action statistics")
add_boolean_option(
group,
"normalize",
"apply normalizations to output in case format is UCCA",
default=False)
ap.add_argument(
"-f",
"--formats",
nargs="+",
choices=FILE_FORMATS,
default=(),
help=
"input formats for creating all parameters before training starts "
"(otherwise created dynamically based on filename suffix), "
"and output formats for written files (each will be written; default: UCCA XML)"
)
ap.add_argument("-u",
"--unlabeled",
nargs="*",
choices=FORMATS,
help="to ignore labels in")
ap.add_argument(
"--lang",
default="en",
help="two-letter language code to use as the default language")
add_boolean_option(
ap, "multilingual",
"separate model parameters per language (passage.attrib['lang'])")
group = ap.add_argument_group(title="Sanity checks")
add_boolean_option(group, "check-loops", "check for parser state loop")
add_boolean_option(group, "verify",
"check for oracle reproducing original passage")
add_boolean_option(group,
"validate-oracle",
"require oracle output to respect constraints",
default=True)
add_param_arguments(ap)
group = ap.add_argument_group(title="DyNet parameters")
group.add_argument("--dynet-mem", help="memory for dynet")
group.add_argument("--dynet-weight-decay",
type=float,
default=1e-5,
help="weight decay for parameters")
add_boolean_option(group,
"dynet-apply-weight-decay-on-load",
"workaround for clab/dynet#1206",
default=False)
add_boolean_option(group, "dynet-gpu", "GPU for training")
group.add_argument("--dynet-gpus",
type=int,
default=1,
help="how many GPUs you want to use")
add_boolean_option(group, "dynet-autobatch",
"auto-batching of training examples")
add_boolean_option(group, "dynet-check-validity",
"check validity of expressions immediately")
DYNET_ARG_NAMES.update(get_group_arg_names(group))
ap.add_argument(
"-H",
"--hyperparams",
type=HyperparamsInitializer.action,
nargs="*",
help=
"shared hyperparameters or hyperparameters for specific formats, "
'e.g., "shared --lstm-layer-dim=100 --lstm-layers=1" "ucca --word-dim=300"',
default=[HyperparamsInitializer.action("shared --lstm-layers 2")])
ap.add_argument("--copy-shared",
nargs="*",
choices=FORMATS,
help="formats whose parameters shall be "
"copied from loaded shared parameters")
self.args = FallbackNamespace(ap.parse_args(args if args else None))
if self.args.config:
print("Loading configuration from '%s'." % self.args.config)
if self.args.passages and self.args.write:
os.makedirs(self.args.outdir, exist_ok=True)
if self.args.models:
if not self.args.log:
self.args.log = self.args.models[0] + ".log"
if self.args.dev and not self.args.devscores:
self.args.devscores = self.args.models[0] + ".dev.csv"
if self.args.passages and not self.args.testscores:
self.args.testscores = self.args.models[0] + ".test.csv"
elif not self.args.log:
self.args.log = "parse.log"
self.sub_configs = [
] # Copies to be stored in Models so that they do not interfere with each other
self._logger = self.format = self.hyperparams = self.iteration_hyperparams = None
self._vocab = {}
self.original_values = {}
self.random = np.random
self.update()
def add_param_arguments(
ap=None,
arg_default=None
): # arguments with possible format-specific parameter values
def add_argument(a, *args, **kwargs):
return a.add_argument(*args, **kwargs)
def add(a, *args, default=None, func=add_argument, **kwargs):
arg = func(a,
*args,
default=default if arg_default is None else arg_default,
**kwargs)
try:
RESTORED_ARGS.add(arg.dest)
except AttributeError:
RESTORED_ARGS.update(get_group_arg_names(arg))
def add_boolean(a, *args, **kwargs):
add(a, *args, func=add_boolean_option, **kwargs)
if not ap:
ap = ArgParser()
group = ap.add_argument_group(title="Node labels")
add(group,
"--max-node-labels",
type=int,
default=0,
help="max number of node labels to allow")
add(group,
"--max-node-categories",
type=int,
default=0,
help="max node categories to allow")
add(group,
"--min-node-label-count",
type=int,
default=2,
help="min number of occurrences for a label")
add_boolean(group, "use-gold-node-labels", "gold node labels when parsing")
add_boolean(group, "wikification",
"use Spotlight to wikify any named node")
add_boolean(group,
"node-labels",
"prediction of node labels, if supported by format",
default=True)
group = ap.add_argument_group(title="Structural constraints")
add_boolean(group, "linkage", "linkage nodes and edges")
add_boolean(group, "implicit", "implicit nodes and edges")
add_boolean(group, "remote", "remote edges", default=True)
add_boolean(group, "constraints", "scheme-specific rules", default=True)
add_boolean(group, "require-connected",
"constraint that output graph must be connected")
add(group,
"--orphan-label",
default="orphan",
help="edge label to use for nodes without parents")
add(group,
"--max-action-ratio",
type=float,
default=100,
help="max action/terminal ratio")
add(group,
"--max-node-ratio",
type=float,
default=10,
help="max node/terminal ratio")
add(group, "--max-height", type=int, default=20, help="max graph height")
group = ap.add_mutually_exclusive_group()
add(group,
"--swap",
choices=(REGULAR, COMPOUND),
default=REGULAR,
help="swap transitions")
add(group,
"--no-swap",
action="store_false",
dest="swap",
help="exclude swap transitions")
add(ap,
"--max-swap",
type=int,
default=15,
help="if compound swap enabled, maximum swap size")
group = ap.add_argument_group(
title="General classifier training parameters")
add(group,
"--learning-rate",
type=float,
help="rate for model weight updates (default: by trainer/1)")
add(group,
"--learning-rate-decay",
type=float,
default=0,
help="learning rate decay per iteration")
add(group,
"--swap-importance",
type=float,
default=1,
help="learning rate factor for Swap")
add(group,
"--max-training-per-format",
type=int,
help="max number of training passages per format per iteration")
add_boolean(group,
"missing-node-features",
"allow node features to be missing if not available",
default=True)
add(group,
"--omit-features",
help="string of feature properties to omit, out of " +
FEATURE_PROPERTIES)
add_boolean(
group, "curriculum",
"sort training passages by action prediction accuracy in previous epoch"
)
group = ap.add_argument_group(title="Perceptron parameters")
add(group,
"--min-update",
type=int,
default=5,
help="minimum #updates for using a feature")
SPARSE_ARG_NAMES.update(get_group_arg_names(group))
group = ap.add_argument_group(title="Neural network parameters")
add(group,
"--word-dim-external",
type=int,
default=300,
help="dimension for external word embeddings")
add(group,
"--word-vectors",
help="file to load external word embeddings from (default: GloVe)")
add(group,
"--vocab",
help=
"file mapping integer ID to word form (to avoid loading spaCy), or '-' to use word form"
)
add_boolean(group,
"update-word-vectors",
"external word vectors in training parameters",
default=True)
add(group,
"--word-dim",
type=int,
default=0,
help="dimension for learned word embeddings")
add(group,
"--lemma-dim",
type=int,
default=200,
help="dimension for lemma embeddings")
add(group,
"--tag-dim",
type=int,
default=20,
help="dimension for fine POS tag embeddings")
add(group,
"--pos-dim",
type=int,
default=20,
help="dimension for coarse/universal POS tag embeddings")
add(group,
"--dep-dim",
type=int,
default=10,
help="dimension for dependency relation embeddings")
add(group,
"--edge-label-dim",
type=int,
default=20,
help="dimension for edge label embeddings")
add(group,
"--node-label-dim",
type=int,
default=0,
help="dimension for node label embeddings")
add(group,
"--node-category-dim",
type=int,
default=0,
help="dimension for node category embeddings")
add(group,
"--punct-dim",
type=int,
default=1,
help="dimension for separator punctuation embeddings")
add(group,
"--action-dim",
type=int,
default=3,
help="dimension for input action type embeddings")
add(group,
"--ner-dim",
type=int,
default=3,
help="dimension for input entity type embeddings")
add(group,
"--shape-dim",
type=int,
default=3,
help="dimension for word shape embeddings")
add(group,
"--prefix-dim",
type=int,
default=2,
help="dimension for word prefix embeddings")
add(group,
"--suffix-dim",
type=int,
default=3,
help="dimension for word suffix embeddings")
add(group,
"--output-dim",
type=int,
default=50,
help="dimension for output action embeddings")
add(group,
"--layer-dim",
type=int,
default=50,
help="dimension for hidden layers")
add(group, "--layers", type=int, default=2, help="number of hidden layers")
add(group,
"--lstm-layer-dim",
type=int,
default=500,
help="dimension for LSTM hidden layers")
add(group,
"--lstm-layers",
type=int,
default=0,
help="number of LSTM hidden layers")
add(group,
"--embedding-layer-dim",
type=int,
default=500,
help="dimension for layers before LSTM")
add(group,
"--embedding-layers",
type=int,
default=1,
help="number of layers before LSTM")
add(group,
"--activation",
choices=ACTIVATIONS,
default=DEFAULT_ACTIVATION,
help="activation function")
add(group,
"--init",
choices=INITIALIZERS,
default=DEFAULT_INITIALIZER,
help="weight initialization")
add(group,
"--minibatch-size",
type=int,
default=200,
help="mini-batch size for optimization")
add(group,
"--optimizer",
choices=TRAINERS,
default=DEFAULT_TRAINER,
help="algorithm for optimization")
add(group,
"--loss",
choices=LOSSES,
default=DEFAULT_LOSS,
help="loss function for training")
add(group,
"--max-words-external",
type=int,
default=250000,
help="max external word vectors to use")
add(group,
"--max-words",
type=int,
default=10000,
help="max number of words to keep embeddings for")
add(group,
"--max-lemmas",
type=int,
default=3000,
help="max number of lemmas to keep embeddings for")
add(group,
"--max-tags",
type=int,
default=100,
help="max number of fine POS tags to keep embeddings for")
add(group,
"--max-pos",
type=int,
default=100,
help="max number of coarse POS tags to keep embeddings for")
add(group,
"--max-deps",
type=int,
default=100,
help="max number of dep labels to keep embeddings for")
add(group,
"--max-edge-labels",
type=int,
default=15,
help="max number of edge labels for embeddings")
add(group,
"--max-puncts",
type=int,
default=5,
help="max number of punctuations for embeddings")
add(group,
"--max-action-types",
type=int,
default=10,
help="max number of action types for embeddings")
add(group,
"--max-action-labels",
type=int,
default=100,
help="max number of action labels to allow")
add(group,
"--max-ner-types",
type=int,
default=18,
help="max number of entity types to allow")
add(group,
"--max-shapes",
type=int,
default=30,
help="max number of word shapes to allow")
add(group,
"--max-prefixes",
type=int,
default=30,
help="max number of 1-character word prefixes to allow")
add(group,
"--max-suffixes",
type=int,
default=500,
help="max number of 3-character word suffixes to allow")
add(group,
"--word-dropout",
type=float,
default=0.2,
help="word dropout parameter")
add(group,
"--word-dropout-external",
type=float,
default=0,
help="word dropout for word vectors")
add(group,
"--lemma-dropout",
type=float,
default=0.2,
help="lemma dropout parameter")
add(group,
"--tag-dropout",
type=float,
default=0.2,
help="fine POS tag dropout parameter")
add(group,
"--pos-dropout",
type=float,
default=0.2,
help="coarse POS tag dropout parameter")
add(group,
"--dep-dropout",
type=float,
default=0.5,
help="dependency label dropout parameter")
add(group,
"--node-label-dropout",
type=float,
default=0.2,
help="node label dropout parameter")
add(group,
"--node-dropout",
type=float,
default=0.1,
help="probability to drop features for a whole node")
add(group,
"--dropout",
type=float,
default=0.4,
help="dropout parameter between layers")
add(group,
"--max-length",
type=int,
default=120,
help="maximum length of input sentence")
add(group,
"--rnn",
choices=["None"] + list(RNNS),
default=DEFAULT_RNN,
help="type of recurrent neural network")
add(group,
"--gated",
type=int,
nargs="?",
default=2,
help="gated input to BiRNN and MLP")
NN_ARG_NAMES.update(get_group_arg_names(group))
return ap
