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):
# 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.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()
}, 'frame')
# 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):
# save and load model
self.model_save_prefix = "zmodel"
self.model_save_suffix_curr = ".curr"
self.model_save_suffix_best = ".best"
self.train_preload_model = "" # preload model name before training?
self.model_load_name = "zmodel.best.m" # (test time) load name
# train
self.accu_batch = 1 # how many fb count as one update
self.min_uidx = 0 # min updated (before early stop)
self.max_uidx = 1000 * 100 # max update
# valid
self.valid_ufreq = 1000 # do valid every this udix
self.valid_first = False # valid once at the very start
self.valid_start_uidx = 0 # do valid >=this
# record best & save cidx
self.record_best_start_cidx = 0 # starting from where to record best
self.save_special_start_cidx = Constants.INT_PRAC_MAX # save special ones starting from what cidx
self.save_special_cfreq = Constants.INT_PRAC_MAX # save (specific name model) every cidx
# lrate schedule
self.lrate = SVConf().direct_update(val=0.001,
which_idx="uidx",
mode="none",
m=0.75,
min_val=0.000001)
self.lrate_warmup_uidx = 0 # linear increasing lrate as warmup for how many uidx
self.lrate_decrease_alpha = 0. # as the one in Transformer, sth like -0.5 (after warmup, uidx**alpha)
def __init__(self):
super().__init__()
self.temperature = SVConf().direct_update(val=1.,
which_idx="eidx",
mode="none",
min_val=0.)
self.cf_merge_mode = "Multinomial" # Multinomial/Geometric
self.pred_argmax = True # select the argmax one!
self.pred_mix_probs = False # weight probs rather than logprobs
def __init__(self):
self.cl_starting_cidx = Constants.INT_PRAC_MAX # starting cl-mode when >= this!
self.cl_update_each = False # whether update after each fb
# --
self.cl_lo0 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=0., max_val=1., idx_scale=50)
self.cl_lo1 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=0., max_val=1., idx_scale=50)
self.cl_lo2 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=0., max_val=1., idx_scale=50)
self.cl_lo3 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=0., max_val=1., idx_scale=50)
self.cl_hi0 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=1., max_val=1., idx_scale=50)
self.cl_hi1 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=1., max_val=1., idx_scale=50)
self.cl_hi2 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=1., max_val=1., idx_scale=50)
self.cl_hi3 = SVConf().direct_update(val=1., which_idx="cidx", mode="linear", b=0., k=1., max_val=1., idx_scale=50)
self.cl_rank_idx = [] # by default, 0,1,2,...
def __init__(self):
super().__init__()
# --
self.temperature = SVConf().direct_update(val=1.,
which_idx="eidx",
mode="none",
min_val=0.)
self.detach_weights = True # detach weights?
self.max_cf = 0.999 # act like label-smoothing
self.cf_trg_rel = False # if so, target at relative (exp(s-max(s))); otherwise, absolute gold prob as trg
self.loss_cf = 0.5
def __init__(self):
super().__init__()
# --
self.loss_weights = []
self.pred_weights = [] # note: not used anymore!!
self.pred_sel_idx = -1 # simply select one (by default the last)
# schedule for weighting!!
self.use_lw_num = 0 # make it brief!
self.lw0 = SVConf().direct_update(val=1.,
which_idx="cidx",
mode="none",
b=0.,
k=0.02,
max_val=1.)
self.lw1 = SVConf().direct_update(val=1.,
which_idx="cidx",
mode="none",
b=0.,
k=0.02,
max_val=1.)
self.lw2 = SVConf().direct_update(val=1.,
which_idx="cidx",
mode="none",
b=0.,
k=0.02,
max_val=1.)
self.lw3 = SVConf().direct_update(val=1.,
which_idx="cidx",
mode="none",
b=0.,
k=0.02,
max_val=1.)
self.lw4 = SVConf().direct_update(val=1.,
which_idx="cidx",
mode="none",
b=0.,
k=0.02,
max_val=1.)
def __init__(self):
super().__init__()
# overall conf (for non-component nodes)
# optimizer
self.main_optim = OptimConf()
# lrate factor
self.main_lrf = SVConf().direct_update(val=1.,
which_idx="eidx",
mode="none",
min_val=0.,
max_val=1.)
# different groups (useful for different lrates)
for ii in range(10): # this should be enough!
setattr(self, f"mpg{ii}", NodeGroupConf()) # model-param-group
# dropouts (general default settings)
self.df_hdrop = 0.1 # hdrop
self.df_gdrop = 0.33 # gdrop (always fixed for recurrent connections)
self.df_idrop = 0.33 # idrop for rnn
self.df_fix_drop = False # fix drop for one run for each dropout
# ema
self.ema_conf = EMAConf()
def __init__(self):
# ==
# top/group-level info (used at outside, put here for convenience)
self.group_name = ""
self.group_files = [
] # List: # -> "input_file" or Dict: sub_name -> "input_file"
self.group_tasks = [] # tasks to perform! note: allow sub-name!
self.group_info = {} # extra info?
self.group_joint = False # join all these into one dataset?
# train (train)
self.group_sample_rate = SVConf().direct_update(
val=1., which_idx="cidx", mode="none") # outside_sample by rate
self.group_sample_alpha = 0. # inside_sample by len(inst)**alpha
# eval (test/dev)
self.group_eval_weight = 1. # weight for final eval
# ==
# (static) io
self.R = ReaderGetterConf()
self.W = WriterGetterConf()
# - paths (further we have default ones for "*_gold", "*_output" if not provided in extras)
self.input_dir = "./" # if needed
self.input_file = ""
self.gold_file = "" # by default the same as input_file
self.output_dir = "./" # if needed
self.output_file = ""
self.output_prefix = "_zout" # default output prefix, full will be "{this}.{wset}.json"
# - special
self.preprocessors = [] # need to slightly modify the data?
# self.approx_prev_next = False # approx. setting of prev & next when loading, note: deprecated
self.presample = 1.0 # (>1=N,<1=Rate) random sample how much at the very beginning, as pre-processing for convenience!
self.presample_shuffle = False # whether shuffle in presample?
self.presample_reverse = False # from back to start (for convenience)
# ==
# runtime
self.convert_conf = ZIConverterConf()
self.batch_conf = ZIBatcherConf()