def main(self, path):
"""
:param path: path to dir containing npy and settings files from the experiment
"""
self.path = path
self.ep = np.load("{}/ep.npy".format(self.path))
#get some info from the setttings file
with open("{}/settings".format(self.path)) as infile:
data_name = None
n_sent = None
for l in infile:
if l.startswith("Name of the corpus file: "):
data_name = l.strip().split(" ")[-1]
elif l.startswith("Number of sentences: "):
n_sent = l.strip().split(" ")[-1]
if data_name is None:
print("Not able to retrieve the dataset name.")
if n_sent is None:
print("Not able to retrieve the number of sentences.")
self.data_name = data_name
self.n_sent = eval(n_sent)
if "tree" or "_rel_" in path:
if "_en_" in path:
self.data = ConllCorpus(self.data_name,
howbig=self.n_sent,
lemmas=False)
elif "_nl_" in path:
self.data = ConllCorpus(self.data_name, howbig=self.n_sent)
else:
self.data = TextCorpus(self.data_name, howbig=self.n_sent)
self.prob_thresh = None
self.n = None # max n of clusters per w
def get_tree(n_inst):
trees = ConllCorpus(file_parsed, howbig=1000000, lemmas=False, eval_spec_rels=self.eval_spec_rel,
dirname=self.dirname, lr=self.lr)
trees.prepare_trees()
# extend instances with trees
assert len(trees.train) == n_inst, "Number of parses not equal to number of classification instances."
c_append = 0
for i in range(n_inst):
# we have a parse:
inst = self.normalize_tree(trees.train[c_append], trees.x_dict, c_append)
c_append += 1
# we don't have a parse:
yield inst
def __init__(self, path, lang, decoding=None, use_wordrep_tree=False):
self.path = path
self.lang = lang
self.decoding = decoding
self.use_wordrep_tree = use_wordrep_tree
if self.decoding is None:
print("Decoding method not specified.")
if self.use_wordrep_tree:
self.decoding = "max-product"
else:
self.decoding = "viterbi"
print("Using default: {}".format(self.decoding))
self.n_states = None
self.n_obs = None
self.n_sent = None
self.n_toks = None
self.corpus_file = None
self.n_states, self.n_obs, self.n_sent, self.n_toks, self.corpus_file = \
self.read_params_from_path()
if self.use_wordrep_tree:
if self.lang == "en":
self.dataset = ConllCorpus("{}".format(self.corpus_file),
howbig=self.n_sent,
lemmas=False)
elif self.lang == "nl":
self.dataset = ConllCorpus("{}".format(self.corpus_file),
howbig=self.n_sent)
else:
self.dataset = TextCorpus("{}".format(self.corpus_file),
howbig=self.n_sent)
self.ner_corpus = None
if self.lang == "nl" and not self.use_wordrep_tree:
self.dev_seq, self.test_seq = self.prepare_seqs_nl_dbg(
self.decoding)
# self.test_seq = self.prepare_seqs_nl(self.decoding)
elif self.lang == "nl" and self.use_wordrep_tree:
self.train_seq, self.dev_seq, self.test_seq = self.prepare_trees_nl(
self.decoding)
# self.test_seq = self.prepare_trees_nl(self.decoding)
elif self.lang == "en" and not self.use_wordrep_tree:
self.dev_seq = self.prepare_seqs_en_dbg(self.decoding)
elif self.lang == "en" and self.use_wordrep_tree:
self.dev_seq = self.prepare_trees_en_dbg(self.decoding)
def get_tree(n_inst):
trees = ConllCorpus(file_parsed, howbig=1000000, lemmas=True, eval_spec_rels=self.eval_spec_rel,
dirname=self.dirname, lr=self.lr)
trees.prepare_trees()
# not every instance has a corresponding tree due to errors in parsing
conll_idx = ConllFilesIndex(files_parsed_path)
conll_idx.create_ids_set()
# extend instances with trees
c_append = 0
for i in range(n_inst):
# we have a parse:
if i + 1 in conll_idx.fileids:
inst = self.normalize_tree(trees.train[c_append], trees.x_dict, c_append)
c_append += 1
# we don't have a parse:
else:
inst = None
yield inst
def get_tree(n_inst):
trees = ConllCorpus(file_parsed,
howbig=1000000,
lemmas=False,
eval_spec_rels=self.eval_spec_rel,
dirname=self.dirname,
lr=self.lr)
trees.prepare_trees()
# extend instances with trees
assert len(
trees.train
) == n_inst, "Number of parses not equal to number of classification instances."
c_append = 0
for i in range(n_inst):
# we have a parse:
inst = self.normalize_tree(trees.train[c_append], trees.x_dict,
c_append)
c_append += 1
# we don't have a parse:
yield inst
def get_tree(n_inst):
trees = ConllCorpus(file_parsed,
howbig=1000000,
lemmas=True,
eval_spec_rels=self.eval_spec_rel,
dirname=self.dirname,
lr=self.lr)
trees.prepare_trees()
self.tree_vocab = trees.x_dict
# not every instance has a corresponding tree due to errors in parsing
conll_idx = ConllFilesIndex(files_parsed_path)
conll_idx.create_ids_set()
# extend instances with trees
c_append = 0
for i in range(n_inst):
# we have a parse:
if i + 1 in conll_idx.fileids:
inst = trees.train[c_append]
c_append += 1
# we don't have a parse:
else:
inst = None
yield inst
import numpy as np
from eval.ner.PrepareHmmRep import read_params_from_path
from hmrtm import HMRTM
from readers.conll_corpus import ConllCorpus
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("-rep", "--rep_path", help="directory containing (hmm) word representations files")
parser.add_argument("--use_lemmas", action='store_true', default=False, help="")
args = parser.parse_args()
path = args.rep_path
posttype_f = "{}posttype_cumul.npy".format(path)
n_states, n_obs, n_sent, n_toks, corpus_file, omit_class_cond, omit_emis_cond = read_params_from_path(path)
lemmas = args.use_lemmas
eval_spec_rel = True
lr = False
params_fixed = (np.load("{}ip.npy".format(path)),
np.load("{}tp.npy".format(path)),
np.load("{}fp.npy".format(path)),
np.load("{}ep.npy".format(path)))
dataset = ConllCorpus("{}".format(corpus_file), howbig=n_sent, lemmas=lemmas, eval_spec_rels=eval_spec_rel,
dirname=path, lr=lr)
dataset.train = dataset.prepare_trees_gen() # generator
h = HMRTM(n_states, n_obs, R=len(dataset.r_dict), params=params_fixed, writeout=False, dirname=path,
omit_class_cond=omit_class_cond, omit_emis_cond=omit_emis_cond)
h.obtain_posttypes_cumul(posttype_f, dataset, n_types=h.M, logger=None)
# obtain model parameters
n_states, n_obs, _, _, _, omit_class_cond, omit_emis_cond = read_params_from_path(path)
lemmas = args.use_lemmas
eval_spec_rel = args.synfunc
lr = False
# load model
params_fixed = (np.load("{}ip.npy".format(path)),
np.load("{}tp.npy".format(path)),
np.load("{}fp.npy".format(path)),
np.load("{}ep.npy".format(path)))
# prepare sents for decoding
sents = ConllCorpus(infile, howbig=1000000, lemmas=lemmas, eval_spec_rels=eval_spec_rel, dirname=path, lr=lr)
sents.prepare_trees()
h = HMRTM(n_states, n_obs, R=len(sents.r_dict), params=params_fixed, writeout=False, dirname=path,
omit_class_cond=omit_class_cond, omit_emis_cond=omit_emis_cond) if eval_spec_rel else \
HMTM(n_states, n_obs, params=params_fixed, writeout=False, dirname=path)
with open(args.outfile, "w") as out:
for tree in sents.train:
# obtain posteriors for all nodes
node_to_rep = h.posterior_decode(tree, cont=True)
# get words
for node in tree.get_nonroots():
out.write(
"{} {}\n".format(sents.x_dict.get_label_name(node.name), nparr_to_str(node_to_rep[node.index])))
out.write("\n")
def __init__(self,
path,
lang,
decoding=None,
use_wordrep_tree=False,
use_wordrep_rel=False,
eval_spec_rel=False,
logger=None,
ignore_rel=None,
lr=False,
use_muc=False):
self.path = path
self.lang = lang
self.decoding = decoding
self.use_wordrep_tree = use_wordrep_tree
self.use_wordrep_rel = use_wordrep_rel
self.eval_spec_rel = eval_spec_rel
self.use_muc = use_muc
if self.decoding is None:
print("Decoding method not specified.")
if self.use_wordrep_tree or self.use_wordrep_rel:
self.decoding = "max-product"
else:
self.decoding = "viterbi"
print("Using default: {}".format(self.decoding))
self.n_states = None
self.n_obs = None
self.n_sent = None
self.n_toks = None
self.corpus_file = None
self.logger = logger
self.n_states, self.n_obs, self.n_sent, self.n_toks, self.corpus_file, self.omit_class_cond, self.omit_emis_cond = \
read_params_from_path(self.path)
if self.logger is not None:
self.logger.debug("Preparing self.dataset")
if self.use_wordrep_tree or self.use_wordrep_rel:
lemmas = False if self.lang == "en" else True
self.dataset = ConllCorpus("{}".format(self.corpus_file),
howbig=self.n_sent,
lemmas=lemmas,
eval_spec_rels=self.eval_spec_rel,
dirname=self.path,
lr=lr)
self.ignore_rel = self.dataset.r_dict.get_label_id(
ignore_rel) if ignore_rel is not None else None
if decoding == "posterior_cont_type":
self.dataset.train = self.dataset.prepare_trees_gen(
) # generator
else:
self.dataset = TextCorpus("{}".format(self.corpus_file),
howbig=self.n_sent)
if decoding == "posterior_cont_type":
self.dataset.prepare_chains()
self.ner_corpus = None
if self.lang == "nl" and not (self.use_wordrep_tree
or self.use_wordrep_rel):
self.train_seq, self.dev_seq, self.test_seq = self.prepare_seqs_nl(
self.decoding)
# self.test_seq = self.prepare_seqs_nl(self.decoding)
elif self.lang == "nl" and (self.use_wordrep_tree
or self.use_wordrep_rel):
self.train_seq, self.dev_seq, self.test_seq = self.prepare_trees_nl(
self.decoding, lr=lr)
# self.test_seq = self.prepare_trees_nl(self.decoding)
elif self.lang == "en" and not (self.use_wordrep_tree
or self.use_wordrep_rel):
self.train_seq, self.dev_seq, self.test_seq, self.muc_seq = self.prepare_seqs_en(
self.decoding)
elif self.lang == "en" and (self.use_wordrep_tree
or self.use_wordrep_rel):
self.train_seq, self.dev_seq, self.test_seq, self.muc_seq = self.prepare_trees_en(
self.decoding, lr=lr)
else:
sys.exit("invalid option in PrepareHmmRep")
self.dataset = None
n_sent += 1
dirname = prepare_dirname(hmm_type=hmm_type,
append_string=append_string,
lang=args.lang,
max_iter=max_iter,
N=start_n_states,
n_sent=n_sent,
alpha=alpha,
minibatch_size=minibatch_size)
if args.tree or args.rel or args.lr:
if args.lang == "en":
dataset = ConllCorpus(args.dataset,
howbig=n_sent,
lemmas=lemmas,
spec_rels=args.rel_spec_en,
dirname=dirname,
lr=args.lr)
elif args.lang == "nl":
dataset = ConllCorpus(args.dataset,
howbig=n_sent,
lemmas=lemmas,
spec_rels=args.rel_spec_nl,
dirname=dirname,
lr=args.lr)
else:
dataset = ConllCorpus(args.dataset,
howbig=n_sent,
lemmas=lemmas,
spec_rels=None,
dirname=dirname,
if args.tree or args.rel or args.lr:
reader = Conll07Reader(args.dataset)
sent = reader.getNext()
while sent:
n_sent += 1
sent = reader.getNext()
else:
for l in line_reader(args.dataset):
n_sent += 1
dirname = prepare_dirname(hmm_type=hmm_type, append_string=append_string, lang=args.lang, max_iter=max_iter,
N=start_n_states, n_sent=n_sent, alpha=alpha, minibatch_size=minibatch_size)
if args.tree or args.rel or args.lr:
if args.lang == "en":
dataset = ConllCorpus(args.dataset, howbig=n_sent, lemmas=lemmas, spec_rels=args.rel_spec_en,
dirname=dirname, lr=args.lr)
elif args.lang == "nl":
dataset = ConllCorpus(args.dataset, howbig=n_sent, lemmas=lemmas, spec_rels=args.rel_spec_nl,
dirname=dirname, lr=args.lr)
else:
dataset = ConllCorpus(args.dataset, howbig=n_sent, lemmas=lemmas, spec_rels=None,
dirname=dirname, lr=args.lr)
n_rels = len(dataset.r_dict)
else:
dataset = TextCorpus(args.dataset, howbig=n_sent)
dataset.prepare_chains()
n_obs = len(dataset.x_dict)
writeout = args.writeout
path = args.rep_path
posttype_f = "{}posttype_cumul.npy".format(path)
n_states, n_obs, n_sent, n_toks, corpus_file, omit_class_cond, omit_emis_cond = read_params_from_path(
path)
lemmas = args.use_lemmas
eval_spec_rel = True
lr = False
params_fixed = (np.load("{}ip.npy".format(path)),
np.load("{}tp.npy".format(path)),
np.load("{}fp.npy".format(path)),
np.load("{}ep.npy".format(path)))
dataset = ConllCorpus("{}".format(corpus_file),
howbig=n_sent,
lemmas=lemmas,
eval_spec_rels=eval_spec_rel,
dirname=path,
lr=lr)
dataset.train = dataset.prepare_trees_gen() # generator
h = HMRTM(n_states,
n_obs,
R=len(dataset.r_dict),
params=params_fixed,
writeout=False,
dirname=path,
omit_class_cond=omit_class_cond,
omit_emis_cond=omit_emis_cond)
h.obtain_posttypes_cumul(posttype_f, dataset, n_types=h.M, logger=None)
def create_vocab(self, dataset, lemmas):
d = ConllCorpus(dataset, howbig=1e10, lemmas=lemmas)
return d.x_dict
