def main():
argparser = configargparse.ArgParser(description=desc)
argparser.add_argument("filenames", nargs="+", help="file names to convert and evaluate")
add_verbose_arg(argparser, help="detailed evaluation output")
add_boolean_option(argparser, "wikification", "Spotlight to wikify any named node (for AMR)")
argparser.add_argument("-o", "--out-dir", help="output directory (if unspecified, files are not written)")
args = argparser.parse_args()
scores = []
for pattern in args.filenames:
filenames = glob(pattern)
if not filenames:
raise IOError("Not found: " + pattern)
for filename in filenames:
print("\rConverting '%s'" % filename, end="")
if args.out_dir or args.verbose:
print(flush=True)
basename, ext = os.path.splitext(os.path.basename(filename))
passage_format = ext.lstrip(".")
converters = CONVERTERS.get(passage_format, CONVERTERS["amr"])
evaluator = EVALUATORS.get(passage_format, EVALUATORS["amr"]).evaluate
with open(filename, encoding="utf-8") as f:
for passage, ref, passage_id in converters[0](f, passage_id=basename, return_original=True):
if args.out_dir:
os.makedirs(args.out_dir, exist_ok=True)
outfile = "%s/%s.xml" % (args.out_dir, passage.ID)
print("Writing '%s'..." % outfile, file=sys.stderr, flush=True)
ioutil.passage2file(passage, outfile)
try:
guessed = converters[1](passage, wikification=args.wikification, use_original=False)
except Exception as e:
raise ValueError("Error converting %s back from %s" % (filename, passage_format)) from e
if args.out_dir:
outfile = "%s/%s%s" % (args.out_dir, passage.ID, ext)
print("Writing '%s'..." % outfile, file=sys.stderr, flush=True)
with open(outfile, "w", encoding="utf-8") as f_out:
print("\n".join(guessed), file=f_out)
try:
s = evaluator(guessed, ref, verbose=args.verbose > 1)
except Exception as e:
raise ValueError("Error evaluating conversion of %s" % filename) from e
scores.append(s)
if args.verbose:
print(passage_id)
s.print()
print()
if args.verbose and len(scores) > 1:
print("Aggregated scores:")
Scores(scores).print()
sys.exit(0)
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()
help="file to write aggregated counts to, in CSV format")
add_boolean_option(argparser,
"unlabeled",
"print unlabeled F1 for individual passages",
short="u")
add_boolean_option(argparser,
"enhanced",
"read enhanced dependencies",
default=True)
add_boolean_option(argparser,
"normalize",
"normalize passages before evaluation",
short="N",
default=True)
add_boolean_option(argparser,
"matching-ids",
"skip passages without a match (by ID)",
short="i")
add_boolean_option(argparser,
"basename",
"force passage ID to be file basename",
short="b")
add_boolean_option(argparser, "units", "print mutual and unique units")
add_boolean_option(argparser, "errors",
"print confusion matrix with error distribution")
group = argparser.add_mutually_exclusive_group()
add_verbose_arg(group, help="detailed evaluation output")
add_boolean_option(group, "quiet", "do not print anything", short="q")
ucca_constructions.add_argument(argparser)
main(argparser.parse_args())
def main(args):
for passages, out_dir, lang in read_specs(args):
scores = []
if not args.verbose:
passages = tqdm(passages, unit=" passages", desc="Parsing " + out_dir)
for passage, parsed in ANNOTATORS[args.parser](passages, lang, args.verbose):
if args.write:
write_passage(parsed, args)
else:
map_labels(parsed, args.label_map)
if args.evaluate:
evaluator = EVALUATORS[args.output_format]
_, converter = CONVERTERS[args.output_format]
if converter is not None:
passage, parsed = map(converter, (passage, parsed))
scores.append(evaluator.evaluate(parsed, passage, verbose=args.verbose > 1))
if scores:
Scores(scores).print()
if __name__ == '__main__':
argparser = argparse.ArgumentParser(description=desc)
add_specs_args(argparser)
argparser.add_argument("--parser", choices=PARSERS, default=SPACY, help="dependency parser to use (default: spacy)")
argparser.add_argument("--output-format", choices=CONVERTERS, help="output file format (default: UCCA)")
add_convert_args(argparser)
argparser.add_argument("-e", "--evaluate", action="store_true", help="evaluate against original passages")
argparser.add_argument("-W", "--no-write", action="store_false", dest="write", help="write parsed passages")
add_verbose_arg(argparser, help="detailed output")
main(argparser.parse_args())
"Cannot specify --extra-normalization without --normalize")
return args
if __name__ == '__main__':
argparser = configargparse.ArgParser(description=desc)
argparser.add_argument("filenames",
nargs="+",
help="file names to convert and evaluate")
argparser.add_argument(
"-f",
"--format",
choices=CONVERTERS,
default="amr",
help="default format (if cannot determine by suffix)")
add_verbose_arg(argparser, help="detailed evaluation output")
argparser.add_argument("--units",
action="store_true",
help="print mutual and unique units")
add_boolean_option(argparser, "wikification",
"Spotlight to wikify any named node (for AMR)")
argparser.add_argument(
"-o",
"--out-dir",
help="output directory (if unspecified, files are not written)")
argparser.add_argument("-n",
"--normalize",
action="store_true",
help="normalize passages before conversion")
argparser.add_argument("-e",
"--extra-normalization",
