def _init_corpus(hconf):
"""Start evaluation; generate folds if needed
:rtype: DataConfig or None
"""
can_skip_folds = fp.exists(hconf.fold_file)
msg_skip_folds = ('Skipping generation of fold files '
'(must have been jumpstarted)')
if hconf.runcfg.stage is None:
# standalone: we always have to load the datapack
# because we'll need it for further stages
mpack = _load_harness_multipack(hconf)
if can_skip_folds:
print(msg_skip_folds, file=sys.stderr)
fold_dict = load_fold_dict(hconf.fold_file)
else:
fold_dict = hconf.create_folds(mpack)
return DataConfig(pack=mpack, folds=fold_dict)
elif hconf.runcfg.stage == ClusterStage.start:
if can_skip_folds:
# if we are just running --start and the fold file already
# exists we can even bail out before reading the datapacks
# because that's all we wanted them for
print(msg_skip_folds, file=sys.stderr)
else:
mpack = _load_harness_multipack(hconf)
hconf.create_folds(mpack)
return None
else:
# any other stage: fold files have already been
# created so we just read them in
return DataConfig(pack=_load_harness_multipack(hconf),
folds=load_fold_dict(hconf.fold_file))
def _init_corpus(hconf):
"""Start evaluation; generate folds if needed
:rtype: DataConfig or None
"""
can_skip_folds = fp.exists(hconf.fold_file)
msg_skip_folds = ('Skipping generation of fold files '
'(must have been jumpstarted)')
if hconf.runcfg.stage is None:
# standalone: we always have to load the datapack
# because we'll need it for further stages
mpack = _load_harness_multipack(hconf)
if can_skip_folds:
print(msg_skip_folds, file=sys.stderr)
fold_dict = load_fold_dict(hconf.fold_file)
else:
fold_dict = hconf.create_folds(mpack)
return DataConfig(pack=mpack, folds=fold_dict)
elif hconf.runcfg.stage == ClusterStage.start:
if can_skip_folds:
# if we are just running --start and the fold file already
# exists we can even bail out before reading the datapacks
# because that's all we wanted them for
print(msg_skip_folds, file=sys.stderr)
else:
mpack = _load_harness_multipack(hconf)
hconf.create_folds(mpack)
return None
else:
# any other stage: fold files have already been
# created so we just read them in
return DataConfig(pack=_load_harness_multipack(hconf),
folds=load_fold_dict(hconf.fold_file))
def _do_corpus(lconf):
"Run evaluation on a corpus"
print(_corpus_banner(lconf), file=sys.stderr)
edus_file = edu_input_path(lconf)
if not os.path.exists(edus_file):
exit_ungathered()
has_stripped = (lconf.stage in [ClusterStage.end, ClusterStage.start]
and fp.exists(features_path(lconf, stripped=True)))
dpack = load_data_pack(edus_file,
pairings_path(lconf),
features_path(lconf, stripped=has_stripped),
verbose=True)
if _is_standalone_or(lconf, ClusterStage.start):
_generate_fold_file(lconf, dpack)
dconf = DataConfig(pack=dpack,
folds=load_fold_dict(lconf.fold_file))
if _is_standalone_or(lconf, ClusterStage.main):
foldset = lconf.folds if lconf.folds is not None\
else frozenset(dconf.folds.values())
for fold in foldset:
_do_fold(lconf, dconf, fold)
if _is_standalone_or(lconf, ClusterStage.combined_models):
mk_combined_models(lconf, dconf)
if _is_standalone_or(lconf, ClusterStage.end):
mk_global_report(lconf, dconf)
def create_folds(self, mpack):
"""
Generate the folds file; return the resulting folds
"""
if FIXED_FOLD_FILE is None:
rng = mk_rng()
fold_dict = make_n_fold(mpack, 10, rng)
else:
fold_dict = load_fold_dict(FIXED_FOLD_FILE)
save_fold_dict(fold_dict, self.fold_file)
return fold_dict
def create_folds(self, mpack):
"""
Generate the folds file; return the resulting folds
"""
if FIXED_FOLD_FILE is None:
rng = mk_rng()
fold_dict = make_n_fold(mpack, 10, rng)
else:
fold_dict = load_fold_dict(FIXED_FOLD_FILE)
save_fold_dict(fold_dict, self.fold_file)
return fold_dict