def postcompute_input1(self,
input0_package,
input1,
mask0=None,
mask1=None,
rel1_t=None):
# ff_score0, ff2_hid0, biaff_hid0 = input0_package
ff_score0, biaff_hid0 = input0_package
ret = 0
if rel1_t is not None:
input1 = input1 + rel1_t
if self.use_ff:
ret = ff_score0 + self.A1(input1)
if self.use_ff2:
zfatal("Not supported in this mode!")
if self.use_biaffine:
# [*, in1, out]
expr0 = biaff_hid0.view(
BK.get_shape(biaff_hid0)[:-1] + [self.in_size1, self.out_size])
# [*, 1, in1] * [*, in1, out] -> [*, 1, out] -> [*, out]
expr1 = BK.matmul(input1.unsqueeze(-2), expr0).squeeze(-2)
ret += expr1 / self.biaffine_div
if self.use_bias:
ret += self.B
# mask
if mask0 is not None:
ret += self.mask_value * (1. - mask0).unsqueeze(-1)
if mask1 is not None:
ret += self.mask_value * (1. - mask1).unsqueeze(-1)
return self.drop_node(ret)
def valid_bigrams(scheme, tag_forms, bos):
types = ChunksSeq.collect_types(tag_forms)
#
rets = {(bos, "O"), ("O", "O")}
if scheme == "BIO":
for t in types:
rets.add((bos, "B-"+t))
rets.add(("O", "B-"+t))
rets.add(("B-"+t, "I-"+t))
rets.add(("I-"+t, "I-"+t))
rets.add(("B-"+t, "O"))
rets.add(("I-"+t, "O"))
for t2 in types:
rets.add(("B-"+t, "B-"+t2))
rets.add(("I-"+t, "B-"+t2))
elif scheme == "BIOES":
for t in types:
rets.add((bos, "B-"+t))
rets.add(("O", "B-"+t))
rets.add((bos, "S-"+t))
rets.add(("O", "S-"+t))
rets.add(("B-"+t, "I-"+t))
rets.add(("I-"+t, "I-"+t))
rets.add(("B-"+t, "E-"+t))
rets.add(("I-"+t, "E-"+t))
rets.add(("S-"+t, "O"))
rets.add(("E-"+t, "O"))
for t2 in types:
rets.add(("S-"+t, "B-"+t2))
rets.add(("E-"+t, "B-"+t2))
rets.add(("S-"+t, "S-"+t2))
rets.add(("E-"+t, "S-"+t2))
else:
zfatal("Unknown tagging scheme")
return rets
def __init__(self, model_type, args):
# top-levels
self.model_type = model_type
self.conf_output = ""
self.log_file = Logger.MAGIC_CODE
self.msp_seed = 9341
#
self.niconf = NIConf() # nn-init-conf
self.dconf = DConf() # data-conf
# parser-conf
if model_type == "simple":
from ..models2.model import MySimpleIEModelConf
self.mconf = MySimpleIEModelConf()
elif model_type == "m3":
from ..models3.model import M3IEModelConf
self.mconf = M3IEModelConf()
elif model_type == "m3a":
from ..models3.modelA import M3AIEModelConf
self.mconf = M3AIEModelConf()
elif model_type == "m3r":
from ..models3.modelR import M3RIEModelConf
self.mconf = M3RIEModelConf()
else:
zfatal(f"Unknown model type: {model_type}, please provide correct type!")
# =====
#
if args is not None:
self.update_from_args(args)
self.validate()
def get_data_writer(file_or_fd, output_format):
if output_format == "conllu":
return ParserConlluWriter(file_or_fd)
elif output_format == "plain":
zwarn("May be meaningless to write plain files for parses!")
return ParserPlainWriter(file_or_fd)
elif output_format == "json":
return ParserJsonWriter(file_or_fd)
else:
zfatal(
"Unknown output_format %s, should select from {conllu,plain,json}"
% output_format)
def precompute_input0(self, input0):
ff_score0 = None
ff2_hid0 = None
biaff_hid0 = None
if self.use_ff:
ff_score0 = self.A0(input0)
if self.use_ff2:
# todo(+3)
zfatal("Not supported in this mode!")
if self.use_biaffine:
biaff_hid0 = BK.matmul(input0, self.W)
# return (ff_score0, ff2_hid0, biaff_hid0)
return (ff_score0, biaff_hid0)
def _marginal(self, full_score_expr, maske_expr, lengths_arr):
if self.alg_unproj:
marginals_expr = nmarginal_unproj(full_score_expr,
maske_expr,
lengths_arr,
labeled=True)
elif self.alg_proj:
marginals_expr = nmarginal_proj(full_score_expr,
maske_expr,
lengths_arr,
labeled=True)
else:
zfatal(
"Unsupported marginal-calculation for the decoding algorithm of "
+ self.conf.iconf.dec_algorithm)
marginals_expr = None
return marginals_expr
def get_data_reader(file_or_fd,
input_format,
use_la0,
noef_link0,
aux_repr_file=None,
max_evt_layers=100):
if input_format == "json":
r = MyDocReader(file_or_fd,
use_la0,
noef_link0,
alter_carg_by_coref=True,
max_evt_layers=max_evt_layers)
else:
zfatal("Unknown input_format %s" % input_format)
r = None
if aux_repr_file is not None and len(aux_repr_file) > 0:
r = AuxDataReader(r, aux_repr_file, "aux_repr")
return r
def build_model(model_type, conf, vpack):
mconf = conf.mconf
model = None
if model_type == "simple":
from ..models2.model import MySimpleIEModel
model = MySimpleIEModel(mconf, vpack)
elif model_type == "m3":
from ..models3.model import M3IEModel
model = M3IEModel(mconf, vpack)
elif model_type == "m3a":
from ..models3.modelA import M3AIEModel
model = M3AIEModel(mconf, vpack)
elif model_type == "m3r":
from ..models3.modelR import M3RIEModel
model = M3RIEModel(mconf, vpack)
else:
zfatal(f"Unknown model type: {model_type}, please provide correct type!")
return model
def __init__(self, partype, args):
# top-levels
self.partype = partype
self.conf_output = ""
self.log_file = Logger.MAGIC_CODE
self.msp_seed = 9341
#
self.niconf = NIConf() # nn-init-conf
self.dconf = DConf() # data-conf
# parser-conf
if partype == "graph":
from ..graph.parser import GraphParserConf
self.pconf = GraphParserConf()
elif partype == "td":
from ..transition.topdown.parser import TdParserConf
self.pconf = TdParserConf()
elif partype == "ef":
from ..ef.parser import EfParserConf
self.pconf = EfParserConf()
elif partype == "g1":
from ..ef.parser import G1ParserConf
self.pconf = G1ParserConf()
elif partype == "g2":
from ..ef.parser import G2ParserConf
self.pconf = G2ParserConf()
elif partype == "s2":
from ..ef.parser import S2ParserConf
self.pconf = S2ParserConf()
elif partype == "fp":
from ..zfp.fp import FpParserConf
self.pconf = FpParserConf()
else:
zfatal(
f"Unknown parser type: {partype}, please provide correct type with the option."
)
# =====
#
self.update_from_args(args)
self.validate()
def output_tags(self, scheme):
ret = []
if scheme == "BIO":
begin_tag, end_tag, single_tag = "B", "I", "B"
elif scheme == "BIOES":
begin_tag, end_tag, single_tag = "B", "E", "S"
else:
zfatal("Unknown tagging scheme")
#
for one in self.all_chunks:
start, end, type = one
if type:
if end-start==1:
ret.append(single_tag+"-"+type)
else:
ret.append(begin_tag+"-"+type)
for _ in range(end-start-2):
ret.append("I-"+type)
ret.append(end_tag+"-"+type)
else:
ret.append("O")
zcheck(len(ret)==self.length, "Err length.")
return ret
def _decode(self, full_score_expr, maske_expr, lengths_arr):
if self.alg_unproj:
return nmst_unproj(full_score_expr,
maske_expr,
lengths_arr,
labeled=True,
ret_arr=True)
elif self.alg_proj:
return nmst_proj(full_score_expr,
maske_expr,
lengths_arr,
labeled=True,
ret_arr=True)
elif self.alg_greedy:
return nmst_greedy(full_score_expr,
maske_expr,
lengths_arr,
labeled=True,
ret_arr=True)
else:
zfatal("Unknown decoding algorithm " +
self.conf.iconf.dec_algorithm)
return None
def build_model(partype, conf, vpack):
pconf = conf.pconf
parser = None
if partype == "graph":
# original first-order graph with various output constraints
from ..graph.parser import GraphParser
parser = GraphParser(pconf, vpack)
elif partype == "td":
# re-implementation of the top-down stack-pointer parser
zwarn(
"Warning: Current implementation of td-mode is deprecated and outdated."
)
from ..transition.topdown.parser import TdParser
parser = TdParser(pconf, vpack)
elif partype == "ef":
# generalized easy-first parser
from ..ef.parser import EfParser
parser = EfParser(pconf, vpack)
elif partype == "g1":
# first-order graph parser
from ..ef.parser import G1Parser
parser = G1Parser(pconf, vpack)
elif partype == "g2":
# higher-order graph parser
from ..ef.parser import G2Parser
parser = G2Parser(pconf, vpack)
elif partype == "s2":
# two-stage parser
from ..ef.parser import S2Parser
parser = S2Parser(pconf, vpack)
elif partype == "fp":
# the finale parser
from ..zfp.fp import FpParser
parser = FpParser(pconf, vpack)
else:
zfatal("Unknown parser type: %s")
return parser
def get_data_reader(file_or_fd,
input_format,
aug_code,
use_la0,
aux_repr_file=None,
aux_score_file=None,
cut=None):
cut = -1 if (cut is None or len(cut) == 0) else int(cut)
if input_format == "conllu":
r = ParseConlluReader(file_or_fd, aug_code, use_la0=use_la0, cut=cut)
elif input_format == "plain":
r = ParseTextReader(file_or_fd, aug_code, cut=cut)
elif input_format == "json":
r = ParseJsonReader(file_or_fd, aug_code, use_la0=use_la0, cut=cut)
else:
zfatal(
"Unknown input_format %s, should select from {conllu,plain,json}" %
input_format)
r = None
if aux_repr_file is not None and len(aux_repr_file) > 0:
r = AuxDataReader(r, aux_repr_file, "aux_repr")
if aux_score_file is not None and len(aux_score_file) > 0:
r = AuxDataReader(r, aux_score_file, "aux_score")
return r
def refresh(self, rop=None):
zfatal("Should call special_refresh instead!")
def get_data_writer(file_or_fd, output_format):
if output_format == "json":
return MyDocWriter(file_or_fd)
else:
zfatal("Unknown output_format %s" % output_format)
def _restart(self):
if self.it_ is None or self.restartable:
self.it_ = iter(self.src_)
else:
zfatal("Cannot restart a non-repeatable stream")
def _restart(self):
if self.started:
zfatal("Cannot restart this fd stream")
self.started = True
def __init__(self, pc: BK.ParamCollection, econf: EncConf):
super().__init__(pc, None, None)
self.conf = econf
#
self.input_dim = econf._input_dim
self.enc_hidden = econf.enc_hidden
# add the sublayers
self.layers = []
# todo(0): allowing repeated names
last_dim = self.input_dim
for name in econf.enc_ordering:
if name == "rnn":
if econf.enc_rnn_layer > 0:
rnn_bidirect, rnn_sep_bidirection = econf.enc_rnn_bidirect, econf.enc_rnn_sep_bidirection
rnn_enc_size = self.enc_hidden // 2 if rnn_bidirect else self.enc_hidden
rnn_layer = self.add_sub_node(
"rnn",
RnnLayerBatchFirstWrapper(
pc,
RnnLayer(pc,
last_dim,
rnn_enc_size,
econf.enc_rnn_layer,
node_type=econf.enc_rnn_type,
bidirection=rnn_bidirect,
sep_bidirection=rnn_sep_bidirection)))
self.layers.append(rnn_layer)
# todo(+2): different i/o sizes for cnn and att?
elif name == "cnn":
if econf.enc_cnn_layer > 0:
per_cnn_size = self.enc_hidden // len(
econf.enc_cnn_windows)
cnn_layer = self.add_sub_node(
"cnn",
Sequential(pc, [
CnnLayer(pc,
last_dim,
per_cnn_size,
econf.enc_cnn_windows,
act="elu")
for _ in range(econf.enc_cnn_layer)
]))
self.layers.append(cnn_layer)
elif name == "att":
if econf.enc_att_layer > 0:
zcheck(last_dim == self.enc_hidden,
"I/O should have same dim for Att-Enc")
att_layer = self.add_sub_node(
"att",
TransformerEncoder(
pc,
econf.enc_att_layer,
last_dim,
econf.enc_att_ff,
econf.enc_att_add_wrapper,
econf.enc_att_conf,
final_act=econf.enc_att_final_act,
fixed_range_vals=econf.enc_att_fixed_ranges))
self.layers.append(att_layer)
elif name == "att2":
if econf.enc_att2_layer > 0:
zcheck(last_dim == self.enc_hidden,
"I/O should have same dim for Att-Enc")
att2_layer = self.add_sub_node(
"att2",
Transformer2Encoder(
pc,
econf.enc_att2_layer,
last_dim,
econf.enc_att2_conf,
short_range=econf.enc_att2_short_range,
long_ranges=econf.enc_att2_long_ranges))
self.layers.append(att2_layer)
else:
zfatal("Unknown encoder name: " + name)
if len(self.layers) > 0:
last_dim = self.layers[-1].get_output_dims()[-1]
self.output_dim = last_dim
#
if econf.no_final_dropout:
self.disable_final_dropout()