def write_tsv(records: Iterator[DialogRecord], path: Union[str, Path]):
seqs = ((str(x), ' '.join(map(str, y))) for x, y in records)
lines = (f'{x}\t{y}\n' for x, y in seqs)
log.info(f"Storing data at {path}")
with IO.writer(path) as f:
for line in lines:
f.write(line)
def store_model(self, step: int, model, train_score: float,
val_score: float, keep: int):
"""
saves model to a given path
:param step: step number of training
:param model: model object itself
:param train_score: score of model on training split
:param val_score: score of model on validation split
:param keep: number of good models to keep, bad models will be deleted
:return:
"""
# TODO: improve this by skipping the model save if the model is not good enough to be saved
if self.read_only:
log.warning("Ignoring the store request; experiment is readonly")
return
name = f'model_{step:05d}_{train_score:.6f}_{val_score:.6f}.pkl'
path = self.model_dir / name
log.info(f"Saving... step={step} to {path}")
torch.save(model, str(path))
for bad_model in self.list_models(sort='total_score',
desc=False)[keep:]:
log.info(f"Deleting bad model {bad_model} . Keep={keep}")
os.remove(str(bad_model))
with IO.writer(os.path.join(self.model_dir, 'scores.tsv'),
append=True) as f:
cols = [
str(step),
datetime.now().isoformat(), name, f'{train_score:g}',
f'{val_score:g}'
]
f.write('\t'.join(cols) + '\n')
def write_lines(path: Union[str, Path], lines):
count = 0
with IO.writer(path) as out:
for line in lines:
count += 1
out.write(line.strip())
out.write("\n")
log.info(f"Wrote {count} lines to {path}")
def write_dialogs(dialogs: Iterator[Dialog], out: Path, dialog_sep='\n'):
count = 0
with IO.writer(out) as outh:
for dialog in dialogs:
count += 1
for utter in dialog.chat:
if utter.uid:
outh.write(f'{utter.uid}\t')
if utter.weight:
outh.write(f'{utter.weight:g}\t')
text = " ".join(map(str, utter.text))
outh.write(f'{utter.char}\t{text}\n')
outh.write(dialog_sep)
log.info(f"Wrote {count} recs to {out}")
def store_config(self):
with IO.writer(self._config_file) as fp:
return yaml.dump(self.config, fp, default_flow_style=False)