def make_feature_maker(algo_name, training_data_dir,
train_set_name, dev_set_name):
logging.info("[%s] Initializing feature maker ...", algo_name)
train_path = os.path.join(training_data_dir, train_set_name + ".csv")
dev_path = os.path.join(training_data_dir, dev_set_name + ".csv")
data_info_path = data_info_lib.data_info_path_for_testing(training_data_dir)
return feature_lib.FeatureMaker(train_path, dev_path, data_info_path)
def main(unused_argv):
if not FLAGS.dev_data_file:
raise ValueError("Specify --dev_data_file")
if not FLAGS.training_data_file:
raise ValueError("Specify --training_data_file")
if not FLAGS.data_info_file:
raise ValueError("Specify --data_info_file")
if not (FLAGS.target_feature or FLAGS.target_feature_file):
raise ValueError("Specify --target_feature or --target_feature_file")
if FLAGS.cross_validate and not FLAGS.best_configurations_file:
raise ValueError(
"Specify --best_configurations_file in cross-validation "
"mode")
features = []
if FLAGS.target_feature_file:
with open(FLAGS.target_feature_file) as s:
for line in s:
features.append(line.strip())
else:
features = [FLAGS.target_feature]
# Process features.
feature_maker = feature_lib.FeatureMaker(FLAGS.training_data_file,
FLAGS.dev_data_file,
FLAGS.data_info_file)
# Perform cross-validation to establish the best configurations of models
# and features or simply train and evaluate.
if FLAGS.cross_validate:
best_configs = _cross_validation_training(feature_maker, features)
logging.info("Saving best configs to \"%s\" ...",
FLAGS.best_configurations_file)
with open(FLAGS.best_configurations_file, "w",
encoding=const.ENCODING) as f:
json.dump(best_configs, f)
else:
_train_and_evaluate(feature_maker, features)
def main(unused_argv):
feature_maker = lib.FeatureMaker(
FLAGS.training_data,
FLAGS.dev_data,
FLAGS.data_info)
training_df, dev_df = feature_maker.process_data(
"Order_of_Subject,_Object_and_Verb")
long_implicational = (
"The_Position_of_Negative_Morphemes_in_SOV_Languages"
"@18 SV&OV&NegV@Order_of_Subject,_Object_and_Verb_majval")
assert "family_majval" in dev_df.columns
assert "family_count" in dev_df.columns
assert long_implicational in dev_df.columns
assert long_implicational in training_df.columns
non_zeroes = []
for fname in training_df.columns:
if ("majval" in fname and
"genus" not in fname and
"family" not in fname and
"neighborhood" not in fname):
for i in training_df[fname]:
if i:
non_zeroes.append(i)
# Show that there are non-zero (non-NA) entries for implicationals:
assert non_zeroes
non_zeroes = []
for fname in dev_df.columns:
if ("majval" in fname and
"genus" not in fname and
"family" not in fname and
"neighborhood" not in fname):
for i in dev_df[fname]:
if i:
non_zeroes.append(i)
# Show that there are non-zero (non-NA) entries for implicationals:
assert non_zeroes
# Remove some of the columns.
#
# Obviously if you use this make sure you do the same thing to both training
# and dev.
smaller_dev_df = feature_maker.select_columns(
dev_df, discard_counts=True, discard_implicationals=True)
assert "wals_code" in smaller_dev_df.columns
assert "target_value" in smaller_dev_df.columns
assert long_implicational not in smaller_dev_df.columns
assert "family_count" not in smaller_dev_df.columns
# Remove some different columns
smaller_dev_df = feature_maker.select_columns(
dev_df, discard_counts=True, discard_implicationals=False)
assert "wals_code" in smaller_dev_df.columns
assert "target_value" in smaller_dev_df.columns
assert long_implicational in smaller_dev_df.columns
assert "family_count" not in smaller_dev_df.columns
# Try another feature
training_df, dev_df = feature_maker.process_data("Hand_and_Arm")
assert "family_majval" in dev_df.columns
assert "family_count" in dev_df.columns
long_implicational = (
"Number_of_Cases@9 Exclusively borderline case-marking"
"@Hand_and_Arm_majval")
assert long_implicational in dev_df.columns
for c in dev_df.columns:
assert c in training_df.columns