def ensemble_errant_result(dataset, stage):
result_path = make_ensemble_base_dir(dataset, stage) / 'result.cat2'
if result_path.exists():
result = ErrantCat2Result(None, None, result_path)
else:
result = None
return result
def write_test(self):
base_dir = make_ensemble_base_dir(self.dataset, 'test')
base_dir.mkdir(parents=True, exist_ok=True)
checkpoint_path = get_ensemble_checkpoint_path(
self.make_result_table())
self.write(base_dir, checkpoint_path, self.test_buffer_size,
self.test_source)
def make_ensemble_result(result_class, dataset, stage):
base_dir = make_ensemble_base_dir(dataset, stage)
result_path = base_dir / 'result.res'
if result_path.exists():
result = result_class(None, None, result_path)
else:
result = None
return result
def test_rescore_errant_result(dataset):
valid_lst = valid_rescore_errant_results(dataset)
if len(valid_lst) > 0:
lmil = int(max(valid_lst).l * 1000)
result_path = make_ensemble_base_dir(dataset, 'test') / 'result.{}.cat2'.format(lmil)
result = ErrantCat2Result(None, None, result_path)
else:
result = None
return result
def show_test_rescore_result(dataset, valid_result_list_fn, test_result_class):
valid_result_list = valid_result_list_fn(dataset, 'valid')
if len(valid_result_list) > 0:
valid_maximum = max(valid_result_list)
best_l = valid_maximum.l
best_lmil = int(best_l * 1000)
result_path = make_ensemble_base_dir(
dataset, 'test') / 'result.{}.res'.format(best_lmil)
test_result = test_result_class(None, None, result_path, l=best_l)
print('rescore (l={}): {}'.format(test_result.l, test_result.gleu))
def valid_rescore_errant_results(dataset):
config = load_config()
result_list = []
for l in config['rescore']['lambda']:
base_dir = make_ensemble_base_dir(dataset, 'valid')
lmil = int(l * 1000)
output = base_dir / 'result.{}.cat2'.format(lmil)
if output.exists():
result = ErrantCat2Result(None, None, output, l = l)
result_list.append(result)
return result_list
def make_rescore_result_list(result_class, dataset, stage):
config = load_config()
result_list = []
for l in config['rescore']['lambda']:
base_dir = make_ensemble_base_dir(dataset, stage)
lmil = int(l * 1000)
output = base_dir / 'result.{}.res'.format(lmil)
if output.exists():
result = result_class(None, None, output, l=l)
result_list.append(result)
return result_list
def __call__(self):
self.write(make_ensemble_base_dir(self.dataset, 'test'))
def make(self):
self.append_command(make_ensemble_base_dir(self.dataset, 'valid'))
self.append_command(make_ensemble_base_dir(self.dataset, 'test'))
def make(self):
for l in self.config['rescore']['lambda']:
lmil = int(l * 1000)
base_dir = make_ensemble_base_dir(self.dataset, self.stage)
code_path = base_dir / 'score.{}.sh'.format(lmil)
self.append_command(base_dir, code_path=code_path)