def __init__(self):
super().__init__()
# --
# the adapter idecs
self.idec_ff = ConfEntryChoices({"yes": IdecConf.make_conf('ff'), "no": None}, "no")
self.idec_satt = ConfEntryChoices({"yes": IdecConf.make_conf('satt'), "no": None}, "no")
# --
# how to deal with msent in training and testing (for some decoders)
self.msent_loss_center = False
self.msent_pred_center = False
def __init__(self):
# vocab dir
self.vocab_save_dir = "./" # place to store all vocabs
self.vocab_load_dir = "" # place to load pre-built vocabs
self.vocab_force_rebuild = False # force rebuild all vocabs
# first is the encoder
self.enc = ConfEntryChoices({
"bert": ZTaskEncBertConf(),
"plain": None
}, "bert")
# then specific tasks -- to be added!!
self.mlm = ConfEntryChoices({"yes": ZTaskMlmConf(), "no": None}, "no")
self.upos = ConfEntryChoices({
"yes": ZTaskUposConf(),
"no": None
}, "no")
self.udep = ConfEntryChoices({
"yes": ZTaskUdepConf(),
"no": None
}, "no")
self.udep2 = ConfEntryChoices(
{
"yes": ZTaskUdepConf().direct_update(name='udep2'),
"no": None
}, "no")
self.pb1 = ConfEntryChoices(
{
"yes": ZTaskSrlConf.make_conf('pb1'),
"no": None
}, "no")
self.pb2 = ConfEntryChoices(
{
"yes": ZTaskSrlConf.make_conf('pb2'),
"no": None
}, "no")
self.pbS = ConfEntryChoices(
{
"yes": ZTaskSrl2Conf.make_conf('pbS'),
"no": None
}, "no")
self.ee = ConfEntryChoices(
{
"yes": ZTaskSrlConf.make_conf('ee'),
"no": None
}, "no")
self.fn = ConfEntryChoices(
{
"yes": ZTaskSrlConf.make_conf('fn'),
"no": None
}, "no")
def __init__(self):
# basic
self.conf_output = "_conf"
self.rgpu = -1
self.cur_run = "1"
self.debug = False
self.do_train = True # training
self.do_test = False # testing
self.do_test_all = False # testing all_df_files
self.train_extras = ""
self.test_extras = ""
# paths
self.run_dir = "" # by default current one!
self.src_dir = "src"
self.voc_dir = "voc"
self.dataset: DataConf = ConfEntryChoices(PRESET_DATA, "pb05")
self.log_prefix = "_log"
self.out_prefix = "_zout"
# --
# specific
self.use_word_input = True # input with word embeddings?
self.use_bert_input = False # input featured bert?
self.use_rel_posi = True # relative position
self.assume_frame = False # assume both trg and frame type
self.assume_trg = False # assume trg
self.no_frame_label = False # no frame label
self.arg_mode = "seq" # seq/span/head
self.arg_seq_mod = "crf" # mle/crf
self.do_ms_train = False # multi-source training?
def make_conf_entry(ftag: str): # make default entries
from msp2.utils import ConfEntryChoices
direct_conf = DirectExtractorConf()
seq_conf = SeqExtractorConf()
soft_conf = SoftExtractorConf()
anchor_conf = AnchorExtractorConf()
all_confs = [direct_conf, seq_conf, soft_conf, anchor_conf]
for cc in all_confs:
cc.ftag = ftag
if ftag == "evt": # simply use shead-mode for evt!!
for cc in all_confs:
cc.core_span_mode = "shead"
cc.lab_conf.labeler_conf.e_dim = 512
# cc.lab_conf.labeler_conf.sconf.pls_conf.hid_dim = 300
# cc.lab_conf.labeler_conf.sconf.pls_conf.hid_nlayer = 1 # give one layer of hidden
cc.lab_conf.labeler_conf.e_tie_weights = True # tie weights for pred&lookup!
direct_conf.span_conf.max_width = 1
elif ftag == "ef" or ftag == "arg":
pass
else:
raise NotImplementedError()
# special for pair-scorer
if ftag == "arg":
for cc in all_confs: # avoid large biaffine-tensor
cc.lab_conf.labeler_conf.sconf.pas_conf.direct_update(
use_biaffine=False, use_ff2=True)
# --
return ConfEntryChoices(
{
"direct": direct_conf,
"seq": seq_conf,
"soft": soft_conf,
"anchor": anchor_conf
}, 'direct')
def __init__(self):
from .model import ZsfpModelConf
from .model2 import MtlModelConf
self.dconf = DConf() # data conf
self.tconf = TConf() # train/test conf
self.mconf = ConfEntryChoices(
{
"basic": ZsfpModelConf(),
"mtl": MtlModelConf()
}, 'basic')
def __init__(self):
# reading and writing formats
# self.R = ReaderGetterConf()
from .run import MyDataReaderConf
self.R = MyDataReaderConf() # reader + optional extra functionalities
self.W = WriterGetterConf()
# --
# data paths: train/dev can be multiple!
self.train: List[str] = []
self.train_props: List[Dict] = [] # extra properties for R
self.dev: List[str] = []
self.test = ""
self.test_do_score = False # do score instead of predict!!
self.cache_data = True # turn off if large data
self.cache_shuffle = True # shuffle cache if using cache
self.dict_dir = "./"
self.dict_frame_file = "" # read frame file for building types
# -----
# for multi-source training (len(train)>1); note: currently three should be enough
self.ms_stop_idx = 0 # with which to consider the end of an epoch
self.ms_train_budget0 = SVConf().direct_update(val=1., which_idx="eidx", mode="none", min_val=0.)
self.ms_train_budget1 = SVConf().direct_update(val=1., which_idx="eidx", mode="none", min_val=0.)
self.ms_train_budget2 = SVConf().direct_update(val=1., which_idx="eidx", mode="none", min_val=0.)
# -----
# save name for trainer not here!!
self.model_load_name = "zmodel.best.m" # load name
self.model_load_strict = True # strict loading?
self.output = "zout.json"
self.eval_conf = ConfEntryChoices(
{'frame': FrameEvalConf(), 'fn': MyFNEvalConf(), 'pb': MyPBEvalConf(), 'dpar': DparEvalConf(), 'none': None}, 'frame')
self.eval2_conf = ConfEntryChoices( # secondary eval
{'frame': FrameEvalConf(), 'fn': MyFNEvalConf(), 'pb': MyPBEvalConf(), 'dpar': DparEvalConf(), 'none': None}, 'none')
# pretrain
self.pretrain_wv_file = ""
self.pretrain_scale = 1.0
# (not used anymore) self.pretrain_init_nohit = 1.0
self.pretrain_hits_outf = "" # output a by-product of filtered embeddings
# extra for testing
self.test_extra_pretrain_wv_files = []
# thresholds for word
self.word_rthres = 200000 # rank <= this
self.word_fthres = 1 # freq >= this
def __init__(self):
super().__init__()
# --
# basic
self.bert_model = "bert-base-multilingual-cased" # or "bert-base-multilingual-cased", "bert-base-cased", "bert-large-cased", "bert-base-chinese", "roberta-base", "roberta-large", "xlm-roberta-base", "xlm-roberta-large", ...
self.bert_cache_dir = "" # dir for downloading
self.bert_no_pretrain = False
self.bert_ft = True # whether fine-tune the model
# --
# extra layers
self.gcn = ConfEntryChoices({'yes': ZGcn0Conf(), 'no': None}, 'no')
def __init__(self):
# node names
self.g_names = []
# optimizer
self.g_optim = ConfEntryChoices({"yes": OptimConf(), "no": None}, "no")
# lrate factor (by default 1.0)
self.g_lrf = SVConf().direct_update(val=1.,
which_idx="eidx",
mode="none",
min_val=0.)
# margin (by default 0.0)
self.g_margin = SVConf().direct_update(val=0.)
def __init__(self):
super().__init__()
# components
# -- input + enc
self.enc_conf = ConfEntryChoices({"plain": ZEncoderPlainConf(), "bert": ZEncoderBertConf()}, "bert")
# -- dec
self.srl_conf = ConfEntryChoices({"srl": ZDecoderSRLConf(), "srl2": ZDecoderSRL2Conf(), "none": None}, "srl2") # srl
self.pos_conf = ConfEntryChoices({"upos": ZDecoderUPOSConf(), "none": None}, "none") # upos
self.dep_conf = ConfEntryChoices({"udep": ZDecoderUDEPConf(), "none": None}, "none") # udep
# --
self.frame_based = False # whether inst should be frame or sent?
# decode
self.decode_sent_thresh_diff = 20 # sent diff thresh in decoding
self.decode_sent_thresh_batch = 8 # num sent one time
# =====
# aug model forward!
self.aug_model_confs = [] # "." means self!!
self.aug_model_paths = [] # should match confs
self.aug_times = 0 # maybe repeat for example, in MC-dropout
self.aug_training_flag = False # when aug forward, training or not?
# =====
# score model forward!
self.score_times = 1 # repeat for scoring, for estimating model confidence.
self.score_training_flag = False # when score forward, training or not?
# =====
# special backward
self.pcgrad = ConfEntryChoices({"yes": PCGradHelperConf(), "none": None}, "none")
self.opt_sparse = ConfEntryChoices({"yes": OptimSparseHelperConf(), "none": None}, "none")
def __init__(self):
super().__init__()
# --
# general specifications
self.evt_ftag = "evt"
self.evt_span_mode = "shead"
self.arg_ftag = "ef" # special target for args
self.arg_span_mode = "span"
self.arg_only_rank1 = True # only consider rank1 arg
# --
# for predicate
self.binary_evt = True # no prediction of detailed lemma+sense?
self.evt_conf = IdecSingleConf() # predicate
self.loss_evt = 1.0
self.evt_label_smoothing = 0.
self.evt_loss_weight_non = 1.0
self.evt_loss_sample_neg = 1.0 # how much of neg to include
# --
# for argument
self.arg_conf: IdecConf = \
ConfEntryChoices({"pair_emb": IdecPairwiseConf(), "pair_att": IdecAttConf()}, "pair_emb") # argument
self.loss_arg = 1.0
self.arg_label_smoothing = 0.
self.arg_loss_sample_neg = 1.0 # how much of neg to include
self.arg_loss_inc_neg_evt = 0. # need to include bad evts? (no args!): a small one is ok
# special for arguments (BIO tagging)
self.arg_use_bio = True # use BIO mode!
self.arg_pred_use_seq_cons = False # use bio constraints in prediction
self.arg_beam_k = 20
# --
# for arg2: aug arg loss
self.arg2_conf: IdecConf = \
ConfEntryChoices({"pair_emb": IdecPairwiseConf(), "pair_att": IdecAttConf()}, "pair_emb") # argument
self.loss_arg2 = 0.0
self.arg2_label_smoothing = 0.
self.arg2_loss_sample_neg = 1.0 # how much of neg to include
# --
# special decoding!
self.evt_pred_use_posi = False # special mode: assume gold posi
self.evt_pred_use_all = False # special debug mode: assume all as targets!
def __init__(self):
super().__init__()
self._isize = -1
# whether do dec's selection?
self.do_dsel = True
self.dsel_conf = DSelectorConf()
self.dsel_nonhit_zero = True # for the nonhit subwords, feed 0, otherwise feed start
# special seq mode
self.do_seq_mode = "none" # none/sel/pool
self.seq_sel_key = "None" # this idx is over original seq (not subword seq)
self.seq_pool = "idx0" # pooling over all items: idx0/avg/max
# --
# extra gate for feeding
_gate_conf = MLPConf().direct_conf(osize=1, dim_hid=256, n_hid_layer=1)
_gate_conf.out_conf.no_drop = True # no drop here
self.gate = ConfEntryChoices({"yes": _gate_conf, "no": None}, "no")
def __init__(self):
super().__init__()
# --
# for depth (distance to root)
self.depth_conf = IdecSingleConf() # one score for depth!
self.depth_label_smoothing = 0.
self.loss_depth = 1.0
self.depth_nonroot_space = 0. # how much space left for nonroot ones? 0 means binary!
# for dep
self.udep_conf: IdecConf = \
ConfEntryChoices({"pair_emb": IdecPairwiseConf(), "pair_att": IdecAttConf()}, "pair_emb") # argument
self.loss_udep = 1.0
self.udep_label_smoothing = 0.
self.udep_loss_sample_neg = 0.05 # how much of neg to include: for each token, (n-1)/n is neg examples
# special
self.udep_pred_root_penalty = -0. # penalty for root edges
self.udep_train_use_cache = True
def __init__(self):
super().__init__()
# --
# udep pairwise label
self.idec_udep_lab = ConfEntryChoices({'idec1': IdecConf.make_conf('pairwise'), 'idec2': Idec2Conf()}, 'idec1')
self.loss_udep_lab = 1.
self.lab_udep = ZLabelConf().direct_update(fixed_nil_val=0., loss_neg_sample=-2.) # 2 neg samples
# special for labels: note: only utilized in training!!
self.udep_ignore_labels = [] # ignore these (by masking out)
self.udep_nilout_labels = [] # make them nil
# root score (simply binary)
self.idec_udep_root = IdecConf.make_conf('score')
self.loss_udep_root = 0. # make it smaller weight
self.lab_root = ZLabelConf().direct_update(fixed_nil_val=0.)
# --
# decoding
self.dec_no_root_score = True # no root score when decoding
self.udep_no_decode = False # simply skip decoding (used in special cases!)
def __init__(self):
super().__init__()
# --
# basic dims
self.ndim = -1
self.nlab = -1
# --
# apply layers
self.app_layers = [] # which layers to apply this idec
self.feed_layers = [
] # which layers to further allow feed (first should have 'app')
# --
# loss & pred
self.hconf: IdecHelperConf = ConfEntryChoices(
{
"simple": IdecHelperSimpleConf(),
"cw": IdecHelperCWConf(),
"cw2": IdecHelperCW2Conf()
}, "simple")
def __init__(self):
super().__init__()
# files
self.take_ratio = 1.0
self.main = ReaderGetterConf() # as gold files
self.gold = ""
self.align_gold_sent = False
self.fake_gold_frames = False # put syntax frames to gold
# eval with preds
self.extra = ReaderGetterConf()
self.preds = [] # list of preds
self.econf: FrameEvalConf = ConfEntryChoices(
{
'frame': FrameEvalConf(),
'fn': MyFNEvalConf(),
'pb': MyPBEvalConf()
}, 'frame')
# self.econf = FrameEvalConf()
# others
self.gold_set_ee_heads = False # try to auto assign heads
self.pred_set_ee_heads = False # try to auto assign heads
def __init__(self):
super().__init__()
# --
self._csize = -1 # number of classes (final output labels)
# todo(+N): not good design, but use this for the indirection ...
self.emb_size = -1 # if >0, we further stack an affine layer for final csize, otherwise directly take inputs!
self.input_act = 'linear' # activation for the input
# fixed_nil
self.fixed_nil_val: float = None # fix NIL as what?
# loss
self.loss_do_sel = False # sel then loss maybe mainly to save some space
self.loss_neg_sample = 1. # how much of neg(NIL) to include in loss, >0 means percentage, <0 means ratio
self.loss_full_alpha = 1. # ordinarily against ALL
# extra binary mode (must set fixed_nil_val=True)
self.use_nil_as_binary = False # use NIL score as binary score
self.loss_binary_alpha = 0. # gold against NIL
# extra all-binary mode (must set nil=0)
self.loss_allbinary_alpha = 0.
# special mode: CRF!
self.crf = ConfEntryChoices({
'yes': ZLinearCrfConf(),
'no': None
}, 'no')
self.pred_crf_add = 100. # todo(+N): simply increase unary scores after decoding
def __init__(self):
super().__init__()
# --
self._isize = -1 # input size (emb)
self._nhead = -1 # input size (num_head)
self._csize = -1 # number of classes
# core scorer
self.cs = CoreScorerConf()
self.share_cs = True
# apply layers
self.app_layers = [] # which layers to apply this idec
self.app_ts = [
] # temp-scales, mainly for prediction!! (by default 1., no ts)
self.app_feeds = [] # whether allow feed (by default 0)
self.app_detach_scales = [
] # gradient scaling at detaching, 0 means full detach! (by default 1.)
self.app_input_fixed_mask = [
] # apply fixed mask at input, 0 means no mask (by default 0.)
# loss & pred
self.hconf: IdecHelperConf = ConfEntryChoices(
{
"simple": IdecHelperSimpleConf(),
"simple2": IdecHelperSimple2Conf()
}, "simple2")
def __init__(self):
super().__init__()
# --
self.conn = ConnectorConf()
self.core = ConfEntryChoices({"ff": CoreFfConf(), "satt": CoreSattConf(),
"score": CoreScorerConf(), "pairwise": CorePairwiseConf()}, "ff")
def __init__(self):
from .model import ZmtlModelConf
self.dconf = DConf() # data conf
self.tconf = TConf() # train/test conf
self.mconf = ConfEntryChoices({"mtl": ZmtlModelConf()}, 'mtl')
