def test_sum(self):
dataset = Dataset('adult')
model = self.fit('Logloss', dataset)
model1 = self.normalize_model(model, dataset, 'test_file')
model2 = self.normalize_model(model, dataset, 'train_file')
s1, b1 = self.get_scale_bias(model1)
s2, b2 = self.get_scale_bias(model2)
# Pick weights w1, w2 so that model_sum = w1 * model1 + w2 * model2
det = (s1 * b2 - s2 * b1)
w1 = b2 / det if det else 0.5
w2 = -b1 / det if det else 0.5
model_sum = get_test_output_path('model_sum{}.bin')
yc.execute([
CATBOOST_PATH,
'model-sum',
'--model-with-weight',
'{}={}'.format(model1, w1),
'--model-with-weight',
'{}={}'.format(model2, w2),
'--output-path',
model_sum,
])
eval_orig = self.eval_model(model, dataset, 'test_file')
eval_sum = self.eval_model(model_sum, dataset, 'test_file')
yc.execute(
get_limited_precision_dsv_diff_tool(1e-8) + [eval_orig, eval_sum])
def test_normalize_idempotent(self):
dataset = Dataset('adult')
model = self.fit('Logloss', dataset)
model_normalized_once = self.normalize_model(model, dataset,
'test_file', 'train_file')
model_normalized_twice = self.normalize_model(model_normalized_once,
dataset, 'test_file',
'train_file')
eval1 = self.eval_model(model_normalized_once, dataset, 'test_file')
eval2 = self.eval_model(model_normalized_twice, dataset, 'test_file')
yc.execute(get_limited_precision_dsv_diff_tool(0) + [eval1, eval2])
def diff_tool(threshold=None):
return get_limited_precision_dsv_diff_tool(threshold, True)
