def main(args):
track = choice_track(args.track)
assert args.subset in track.subsets
assert bool(args.ckpt_dir) ^ bool(args.ckpt_fpath)
if args.find_best:
assert bool(args.ckpt_dir)
databin_path = track.get_databin_path('pretrain')
gold_m2, ori_path, ori_bpe_path, gen_subset, scorer_type = track.get_subset_datapath(
args.subset)
# ckpt_dir
if args.ckpt_dir is not None:
ckpt_files = util.get_sorted_ckpts(args.ckpt_dir)
output_dir = track.get_output_dir(args.ckpt_dir)
# ckpt_fpath
else:
ckpt_files = [args.ckpt_fpath]
output_dir = track.get_output_dir(args.ckpt_fpath)
if not args.find_best:
for ckpt in tqdm(ckpt_files):
run_ckpt(databin_path, ckpt, output_dir, scorer_type, gold_m2,
ori_path, ori_bpe_path, gen_subset, args.remove_unk_edits,
args.remove_error_type_lst, args.apply_rerank,
args.preserve_spell, args.max_edits)
logging.info(f"[Evaluate] highest score on {ori_path}")
find_best(output_dir, ori_path, scorer_type)
def main(args):
track = choice_track(args.track)
prepare_text(track)
for train_mode in track.train_modes:
databin_path = track.get_databin_path(train_mode)
trainpref, validpref = track.get_pref(train_mode)
prepare_binary(databin_path, trainpref, validpref, track.fp.BPE_VOCAB)
def main(args):
track = choice_track(args.track)
assert args.train_mode in track.train_modes
if args.train_mode == 'pretrain':
assert args.prev_model_output_dir is None
databin_path = track.get_databin_path(args.train_mode)
model_config = track.get_model_config(args.model, args.lr, args.dropout, args.max_epoch, args.seed, args.reset)
ckpt_dir = track.get_ckpt_dir(args.train_mode, args.model, args.lr, args.dropout, args.seed,
args.prev_model_output_dir)
_, ori_path, _, _, scorer_type = track.get_subset_datapath('valid')
fscore, restore_ckpt = find_restore(args.prev_model_output_dir, ori_path, scorer_type)
train(databin_path, model_config, ckpt_dir, restore_ckpt, args.ngpu)
def main(args):
# (NOTE)
# 1: Restricted
# 3: Low resource
# 0: CONLL
track = choice_track(args.track)
print_log("------ ------ ------")
print_log(
f"[Prepare] 1. prepare for the text data of track {track.TRACK_NUM}"
) # (MODIFIED)
prepare_text(track)
print_log("------ ------ ------")
print_log(f"[Prepare] 2. create binary data")
for train_mode in track.train_modes:
databin_path = track.get_databin_path(train_mode)
trainpref, validpref = track.get_pref(train_mode)
prepare_binary(databin_path, trainpref, validpref, track.fp.BPE_VOCAB)
def main(args):
track = choice_track(args.track)
assert args.train_mode in track.train_modes
if args.train_mode == 'pretrain':
assert args.prev_model_output_dir is None
# (NOTE) Data used for training.
databin_path = track.get_databin_path(args.train_mode)
# (NOTE) Config.
model_config = track.get_model_config(args.model, args.lr, args.dropout, args.max_epoch, args.seed, args.reset)
# (NOTE) NAME of the checkpoint dir.
ckpt_dir = track.get_ckpt_dir(args.train_mode, args.model, args.lr, args.dropout, args.seed,
args.prev_model_output_dir)
# (NOTE) Path of the checkpoint to be restored.
_, ori_path, _, _, scorer_type = track.get_subset_datapath('valid')
fscore, restore_ckpt = find_restore(args.prev_model_output_dir, ori_path, scorer_type)
train(databin_path, model_config, ckpt_dir, restore_ckpt, args.ngpu)
def main(args):
track = choice_track(args.track)
assert args.subset in track.subsets
assert bool(args.ckpt_dir) ^ bool(args.ckpt_fpath)
if args.find_best:
assert bool(args.ckpt_dir)
databin_path = track.get_databin_path('pretrain')
# (NOTE) gold_m2: trg.
# (NOTE) ori_path: src.
gold_m2, ori_path, ori_bpe_path, gen_subset, scorer_type = track.get_subset_datapath(
args.subset)
# ckpt_dir
if args.ckpt_dir is not None:
# (NOTE) checkpoint_best.pt and checkpoint_last.pt are not included.
ckpt_files = util.get_sorted_ckpts(args.ckpt_dir)
# (NOTE) Path of the eval rsts to be stored.
output_dir = track.get_output_dir(args.ckpt_dir)
# ckpt_fpath
else:
# (NOTE) checkpoint_best.pt and checkpoint_last.pt are not included.
ckpt_files = [args.ckpt_fpath]
# (NOTE) Path of the eval rsts to be stored.
output_dir = track.get_output_dir(args.ckpt_fpath)
if not args.find_best:
# (NOTE) Evals all checkpoints for `ckpt_dir` (checkpoint1.pt, checkpoint2.pt, ... ),
# (NOTE) or the checkpoint file for `ckpt_fpath`.
for ckpt in tqdm(ckpt_files):
run_ckpt(databin_path, ckpt, output_dir, scorer_type, gold_m2,
ori_path, ori_bpe_path, gen_subset, args.remove_unk_edits,
args.remove_error_type_lst, args.apply_rerank,
args.preserve_spell, args.max_edits)
logging.info(f"[Evaluate] highest score on {ori_path}")
# (NOTE) Finds the best fscore and checkpoint from all the *.report.
find_best(output_dir, ori_path, scorer_type)