def __init__(self,
path: str,
dir_file_f: Callable = (lambda x: x.id),
dir_file_suffix=".json",
end="\n"):
super().__init__()
self.path = path
self.dir_file_f = dir_file_f
self.dir_file_suffix = dir_file_suffix
self.end = end
mkdir_p(path, True) # mkdir
def execute(self, env_prefix: str):
conf = self.conf
# first change path if needed!
RUN_DIR = conf.run_dir
if RUN_DIR:
mkdir_p(RUN_DIR, raise_error=True)
# os.chdir(RUN_DIR) # change to it!!
else:
conf.run_dir = "."
# directly change conf
conf.src_dir = zglob1(conf.src_dir, check_prefix="..", check_iter=10)
conf.src_dir = os.path.abspath(conf.src_dir) # get absolute path!!
conf._env_prefix = env_prefix
# --
self.run()
res = self.get_result()
return res
def save(self, prefix="./"):
fname = prefix + "zsfp.voc.pkl"
mkdir_p(os.path.dirname(fname)) # try to make dir if not there!
default_pickle_serializer.to_file([self.vocabs, self.embeds], fname)
def main(*args):
conf: RunConf = init_everything(RunConf(),
args,
add_utils=False,
add_nn=False)
# =====
# get paths
RUN_DIR = conf.run_dir
if RUN_DIR:
mkdir_p(RUN_DIR, raise_error=True)
os.chdir(RUN_DIR) # change to it!!
SRC_DIR = zglob1(conf.src_dir, check_prefix="..", check_iter=10)
VOC_DIR = zglob1(conf.voc_dir, check_prefix="..", check_iter=10)
DATA_DIR = zglob1(conf.dataset.data_dir, check_prefix="..", check_iter=10)
zlog(
f"RUN with RUN={RUN_DIR}, SRC={SRC_DIR}, VOC={VOC_DIR}, DATA={DATA_DIR}"
)
# =====
# modes
dataset_choice = conf.dataset._choice
is_pb, is_fn = [dataset_choice.startswith(z) for z in ["pb", "fn"]]
assert is_pb or is_fn
# =====
# options
# --
# base ones
base_opt = "conf_output:_conf"
# eval
if is_pb:
base_opt += f" eval_conf:pb"
elif is_fn:
base_opt += f" dict_frame_file:{DATA_DIR}/{conf.dataset.frame_file}"
base_opt += f" eval_conf:fn eval_conf.frame_file:{DATA_DIR}/{conf.dataset.frame_file}" # eval
# --
# =====
# modeling
if conf.use_word_input:
base_opt += " ec_word.dim:300 ec_word.drop_rate:0.2 ec_word.init_from_pretrain:1 ec_word.rare_unk_thr:2" # word
# base_opt += " ec_posi.dim:512" # posi?
# base_opt += " ec_char.dim:50 ec_char.init_scale:5." # char?
if conf.use_bert_input:
base_opt += " ec_bert.dim:768 bert_model:bert-base-cased bert_output_layers:7,8,9" # bert features?
base_opt += " eproj_dim:512" # --
if conf.use_rel_posi:
base_opt += " enc_conf.enc_att.n_layers:2 enc_conf.enc_att.use_posi:0 enc_conf.clip_dist:16" # enc1
else:
base_opt += " enc_conf.enc_att.n_layers:2 enc_conf.enc_att.use_posi:1 enc_conf.clip_dist:0" # enc1
# base_opt += " enc_conf.enc_tatt.n_layers:2 enc_conf.enc_tatt.use_posi:1" # enc1
# base_opt += " enc_conf.enc_rnn.n_layers:1 enc_conf.enc_hidden:1024" # enc1
# --
# frame
base_opt += " loss_evt:0.5 pred_evt:1" # with evts
base_opt += " evt_conf.cand_label_smoothing:0.05 evt_conf.label_smoothing:0.1" # label smooth
base_opt += " evt_conf.lookup_conf.use_emb:0" # no adding frame embeddings?
base_opt += " evt_conf.span_conf.sconf.hid_nlayer:1" # pred scorer?
if conf.assume_frame: # no need for the evt module!!
base_opt += " loss_evt:0 pred_evt:0 eval_conf.weight_frame:0."
elif conf.assume_trg: # no need for cand, but still need to identify frame types
base_opt += " evt_conf.loss_cand:0. evt_conf.loss_use_posi:1 evt_conf.pred_use_posi:1" # use-posi for evt
base_opt += " evt_conf.pred_addition_non_score:-100000." # NEGINF-non
if is_fn: # further use cons for fn
base_opt += f" evt_cons_lex_file:{VOC_DIR}/cons_lex_{dataset_choice}.json evt_conf.pred_use_cons:1 evt_conf.pred_use_lu:1 evt_conf.loss_use_cons:0 evt_conf.loss_use_lu:0" # cons & use-lu for evt
else:
# evt_conf -> direct
base_opt += " evt_conf.loss_cand:1.0 evt_conf.loss_lab:1.0" # loss_cand
base_opt += " evt_conf.span_train_sample_rate:1.0 evt_conf.span_topk_rate:1.0 evt_conf.span_train_sample:1" # some rates
# --
if is_pb: # lab is aux for pb
base_opt += " evt_conf.loss_lab:0.5 evt_conf.pred_score_prune:0. evt_conf.pred_addition_non_score:-100000."
elif is_fn: # lab is essential for fn
base_opt += " loss_evt:1 evt_conf.loss_cand:0.5 evt_conf.span_train_sample_rate:0.33 evt_conf.span_topk_rate:0.4 evt_conf.span_train_sample:1"
# --
if conf.no_frame_label:
base_opt += " evt_conf.loss_lab:0. evt_conf.pred_score_prune:0. evt_conf.pred_addition_non_score:-100000."
# --
# arg
base_opt += " arg_use_finput:0"
base_opt += f" fenc_conf.enc_att.n_layers:8 fenc_conf.clip_dist:{16 if conf.use_rel_posi else 0}" # fenc
# base_opt += " fenc_conf.enc_tatt.n_layers:6" # fenc
# base_opt += " fenc_conf.enc_rnn.n_layers:3 fenc_conf.enc_hidden:1024" # enc1
base_opt += " loss_arg:1. pred_arg:1" # with args
base_opt += " arg_conf.label_smoothing:0.1" # label smooth
if conf.arg_mode in ["span", "head"]:
# arg_conf -> direct
base_opt += " arg_conf.loss_cand:0.5" # loss_cand
# base_opt+=" arg_conf.span_train_sample_rate:0.33 arg_conf.span_topk_rate:0.4" # some rates
base_opt += " arg_conf.span_topk_rate:1. arg_conf.span_topk_count:10 arg_conf.span_train_sample:0" # some rates
base_opt += " arg_conf.loss_weight_non:1." # less penalizing this?
base_opt += " arg_conf.pp_check_more:1" # check non-overlapping
if conf.arg_mode == "span":
base_opt += " arg_conf.max_width:30 arg_conf.softhead_topk:5 arg_conf.pred_non_overlapping:1" # span
elif conf.arg_mode == "head":
base_opt += " arg_conf.core_span_mode:shead arg_conf.max_width:1" # head
# extender
base_opt += " arg_conf.loss_ext:0.5 arg_conf.pred_ext:1 arg_conf.ext_use_finput:0"
base_opt += f" arg_conf.ext_conf.eenc_conf.enc_att.n_layers:1 arg_conf.ext_conf.eenc_conf.enc_att.aconf.clip_dist:{16 if conf.use_rel_posi else 0}"
else:
raise NotImplementedError()
elif conf.arg_mode == "soft":
base_opt += " arg_conf:soft"
base_opt += " arg_conf.loss_ext:0.5 arg_conf.pred_ext:1 arg_conf.ext_use_finput:0"
base_opt += f" arg_conf.ext_conf.eenc_conf.enc_att.n_layers:1 arg_conf.ext_conf.eenc_conf.enc_att.aconf.clip_dist:{16 if conf.use_rel_posi else 0}"
base_opt += " arg_conf.pp_check_more:1"
elif conf.arg_mode in ["anchor", "anchor2"]:
base_opt += " arg_conf:anchor"
if conf.arg_mode == "anchor2": # yet another head mode!
base_opt += " arg_conf.core_span_mode:shead"
base_opt += " arg_conf.loss_ext:0.5 arg_conf.pred_ext:1 arg_conf.ext_use_finput:0"
base_opt += f" arg_conf.ext_conf.eenc_conf.enc_att.n_layers:1 arg_conf.ext_conf.eenc_conf.enc_att.aconf.clip_dist:{16 if conf.use_rel_posi else 0}"
base_opt += " arg_conf.pp_check_more:1"
elif conf.arg_mode in ["seq", "seq0"]:
# arg_conf -> seq
base_opt += " arg_conf:seq arg_conf.seq_scheme:BIO" # use seq mode!
base_opt += " arg_conf.loss_weight_non:1." # less penalizing this?
# --
if conf.arg_mode == "seq":
base_opt += " arg_conf.beam_k:150 arg_conf.use_bigram:0 arg_conf.pred_use_seq_cons:1" # viterbi with constraints
if conf.arg_seq_mod == "crf": # crf-mode
base_opt += " arg_conf.loss_mode:crf arg_conf.use_bigram:1 arg_conf.local_normalize:0"
elif conf.arg_mode == "seq0": # greedy mode: no crf and no viterbi
base_opt += " arg_conf.pred_use_seq_cons:0 arg_conf.loss_mode:mle arg_conf.use_bigram:0 arg_conf.local_normalize:1"
else:
raise NotImplementedError()
else:
raise NotImplementedError()
# --
# =====
# training
base_opt += " ema_decay:0. ema_update_step:1" # ema
if 1:
UPE = 1000 # 1000 update as one epoch
base_opt += " lrate.val:0.0002 anneal_times:10 anneal_patience:10 lrate.m:0.75"
base_opt += f" valid_ufreq:{UPE} valid_epoch:0 max_uidx:{UPE*150} lrate_warmup_uidx:{8*UPE} lrate_decrease_alpha:0."
if conf.use_rel_posi:
base_opt += " train_count_mode:ftok train_batch_size:4096 accu_batch:1" # actually bs=bs*accu
base_opt += " test_count_mode:ftok test_batch_size:2048"
else:
base_opt += " train_count_mode:ftok train_batch_size:4096 accu_batch:1" # actually bs=bs*accu
base_opt += " test_count_mode:ftok test_batch_size:2048"
base_opt += " df_hdrop:0.2" # general dropout
else: # possibly for rnn
base_opt += " lrate.val:0.002 anneal_times:10 anneal_patience:10"
base_opt += " train_count_mode:frame max_eidx:100 train_batch_size:32"
base_opt += " df_hdrop:0.33" # general dropout
if is_pb:
base_opt += " train_skip_noevt_rate:0.0"
elif is_fn:
base_opt += " train_skip_noevt_rate:1.0" # skip sents where no targets!
# data
base_opt += " " + conf.dataset.get_data_str(DATA_DIR, conf.do_ms_train)
base_opt += f" pretrain_wv_file:{VOC_DIR}/hits_{dataset_choice}.vec pretrain_scale:10." # filtered pretrain file
# nn
base_opt += f" nn.device:0 nn.random_seed:9347{conf.cur_run} nn.random_cuda_seed:9349{conf.cur_run}"
# =====
# note: base_opt is only for training!!
_L_PRE = conf.log_prefix
DEBUG_OPTION = "-m pdb" if conf.debug else ""
TRAIN_CMD = f"CUDA_VISIBLE_DEVICES={conf.rgpu} PYTHONPATH={SRC_DIR}:$PYTHONPATH python3 {DEBUG_OPTION} -m msp2.tasks.zsfp.main.train {base_opt} log_file:{_L_PRE}_train {conf.train_extras}"
# --
TEST_CMD = f"CUDA_VISIBLE_DEVICES={conf.rgpu} PYTHONPATH={SRC_DIR}:$PYTHONPATH python3 {DEBUG_OPTION} -m msp2.tasks.zsfp.main.test {conf.conf_output} log_file:{_L_PRE}_test {conf.test_extras}"
# --
if conf.do_train:
system(TRAIN_CMD, pp=True)
# --
if conf.do_test:
system(TEST_CMD, pp=True)
# --
if conf.do_test_all:
for tfile in conf.dataset.all_dt_files:
_TMP_CMD = f"CUDA_VISIBLE_DEVICES={conf.rgpu} PYTHONPATH={SRC_DIR}:$PYTHONPATH python3 {DEBUG_OPTION} -m msp2.tasks.zsfp.main.test {conf.conf_output} test:{DATA_DIR}/{tfile} output:{conf.out_prefix}.{tfile} log_file:{_L_PRE}.{tfile} test_extra_pretrain_wv_files:{VOC_DIR}/hits_{dataset_choice}.vec {conf.test_extras}"
system(_TMP_CMD, pp=True)