def main():
args = create_argparser().parse_args()
base_props, props_to_evaluate = get_props_to_evaluate(args)
dataset, unlabeled_docs = load_docs(args)
props_picker, props_best_results = get_best_model_picker(
history_improvement_controller(len(props_to_evaluate)))
if not props_to_evaluate:
print("No props found")
return
if os.path.exists(args.out_dir) and (not os.path.isdir(args.out_dir)
or os.listdir(args.out_dir)):
print("Output path should either not exists or be empty directory")
return
for props_idx, props in enumerate(props_to_evaluate):
cur_props_path = join(args.out_dir, f'props_{props_idx}')
os.makedirs(cur_props_path, exist_ok=True)
dump_dict_as_json(props, join(cur_props_path, 'props.json'))
with transformer_from_props(props.get("transformers", {})) as t:
tr_dataset = dataset.transformed_by(t)
tr_unlabeled_docs = [t.transform(doc) for doc in unlabeled_docs
] if unlabeled_docs is not None else None
mean_main_score, mean_scores = splits_cycle(args.task_name, args.seeds,
props, props_idx,
tr_dataset,
tr_unlabeled_docs,
cur_props_path)
props_picker(mean_main_score, mean_scores, lambda: None)
best_props_idx = props_best_results.best_score_idx
best_main_score, best_scores = props_best_results.best_scores
print(
f"Overall experiment best score: {best_main_score:.4f}, props: #{best_props_idx}"
)
best_props_path = join(args.out_dir, 'best_props')
os.makedirs(best_props_path)
for split_idx in range(dataset.splits_number):
split_path = join(args.out_dir, f'props_{best_props_idx}',
f'split_{split_idx}')
split_best_seed = join(split_path, 'best_seed')
shutil.copytree(split_best_seed,
join(best_props_path, f'split_{split_idx}'))
shutil.copy(join(split_path, 'mean_results.json'),
join(best_props_path, f'split_{split_idx}'))
dump_dict_as_json({
**best_scores, "props_num": best_props_idx
}, join(args.out_dir, "best_results.json"))
dump_dict_as_json(
get_experiments_report(props_best_results, base_props,
props_to_evaluate),
join(args.out_dir, "experiments_report.json"))
def get_evaluating_hook(dev_docs: List[Document], train_docs: List[Document],
evaluate, base_path: str, early_stopping_rounds: int):
stats_path = join(base_path, "best_model_stats")
os.makedirs(stats_path, exist_ok=True)
model_path = join(base_path, "best_model")
os.makedirs(model_path, exist_ok=True)
picker, results_storage = get_best_model_picker(
history_improvement_controller(early_stopping_rounds))
def save_clf_and_dump_stats(classifier, stats_generator):
classifier.save(model_path)
if stats_generator is not None:
for i, doc in enumerate(dev_docs):
with open(join(stats_path, doc.name + '_stats.txt'),
'w',
encoding='utf-8') as f:
f.write(stats_generator(i))
def apply(classifier, epoch):
print(f"Epoch {epoch}, dev results:")
dev_main_score, dev_scores, stats_generator = evaluate(classifier,
dev_docs,
need_stats=True)
print("Score={:.4f}".format(dev_main_score))
print()
stopped = picker(
dev_main_score, dev_scores,
lambda: save_clf_and_dump_stats(classifier, stats_generator))
if epoch % 10 == 0:
print(f"Epoch {epoch}, train results:")
train_main_score, train_scores, _ = evaluate(classifier,
train_docs,
need_stats=False)
print("Score={:.4f}".format(train_main_score))
print()
else:
train_scores = None
scores = {'dev': dev_scores}
if train_scores is not None:
scores['train'] = train_scores
with open(join(base_path, 'results.txt'), 'a', encoding='utf-8') as f:
f.write(
json.dumps({f"epoch_{epoch}": scores},
indent=4,
sort_keys=True))
f.write('\n\n')
return stopped
return apply, results_storage