Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[py] eta decay bugfix

Browse Source
This commit is contained in:
Faron 2016-04-30 12:32:49 +02:00
parent 9bc2ac4bd0
commit ad3f49e881
2 changed files with 33 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
python-package/xgboost/training.py
View File

@ -108,6 +108,7 @@ def train(params, dtrain, num_boost_round=10, evals=(), obj=None, feval=None,
bst = Booster(params, [dtrain] + [d[0] for d in evals]) bst = Booster(params, [dtrain] + [d[0] for d in evals])
_params = dict(params) if isinstance(params, list) else params _params = dict(params) if isinstance(params, list) else params
_eta_param_name = 'eta' if 'eta' in _params else 'learning_rate'
if 'num_parallel_tree' in _params: if 'num_parallel_tree' in _params:
num_parallel_tree = _params['num_parallel_tree'] num_parallel_tree = _params['num_parallel_tree']
nboost //= num_parallel_tree nboost //= num_parallel_tree
@ -168,9 +169,9 @@ def train(params, dtrain, num_boost_round=10, evals=(), obj=None, feval=None,
for i in range(start_iteration, num_boost_round): for i in range(start_iteration, num_boost_round):
if learning_rates is not None: if learning_rates is not None:
if isinstance(learning_rates, list): if isinstance(learning_rates, list):
bst.set_param({'eta': learning_rates[i]}) bst.set_param(_eta_param_name, learning_rates[i])
else: else:
bst.set_param({'eta': learning_rates(i, num_boost_round)}) bst.set_param(_eta_param_name, learning_rates(i, num_boost_round))
# Distributed code: need to resume to this point. # Distributed code: need to resume to this point.
# Skip the first update if it is a recovery step. # Skip the first update if it is a recovery step.

34
tests/python/test_basic_models.py
View File

@ -10,7 +10,6 @@ rng = np.random.RandomState(1994)
class TestModels(unittest.TestCase): class TestModels(unittest.TestCase):
def test_glm(self): def test_glm(self):
param = {'silent': 1, 'objective': 'binary:logistic', param = {'silent': 1, 'objective': 'binary:logistic',
'booster': 'gblinear', 'alpha': 0.0001, 'lambda': 1} 'booster': 'gblinear', 'alpha': 0.0001, 'lambda': 1}
@ -25,12 +24,39 @@ class TestModels(unittest.TestCase):
assert err < 0.1 assert err < 0.1
def test_eta_decay(self): def test_eta_decay(self):
param = {'max_depth': 2, 'eta': 1, 'silent': 1, 'objective': 'binary:logistic'}
watchlist = [(dtest, 'eval'), (dtrain, 'train')] watchlist = [(dtest, 'eval'), (dtrain, 'train')]
num_round = 2 num_round = 4
# learning_rates as a list # learning_rates as a list
bst = xgb.train(param, dtrain, num_round, watchlist, learning_rates=[0.4, 0.3]) # init eta with 0 to check whether learning_rates work
param = {'max_depth': 2, 'eta': 0, 'silent': 1, 'objective': 'binary:logistic'}
evals_result = {}
bst = xgb.train(param, dtrain, num_round, watchlist, learning_rates=[0.8, 0.7, 0.6, 0.5],
evals_result=evals_result)
eval_errors = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster) assert isinstance(bst, xgb.core.Booster)
# validation error should decrease, if eta > 0
assert eval_errors[0] > eval_errors[-1]
# init learning_rate with 0 to check whether learning_rates work
param = {'max_depth': 2, 'learning_rate': 0, 'silent': 1, 'objective': 'binary:logistic'}
evals_result = {}
bst = xgb.train(param, dtrain, num_round, watchlist, learning_rates=[0.8, 0.7, 0.6, 0.5],
evals_result=evals_result)
eval_errors = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
# validation error should decrease, if learning_rate > 0
assert eval_errors[0] > eval_errors[-1]
# check if learning_rates override default value of eta/learning_rate
param = {'max_depth': 2, 'silent': 1, 'objective': 'binary:logistic'}
evals_result = {}
bst = xgb.train(param, dtrain, num_round, watchlist, learning_rates=[0, 0, 0, 0],
evals_result=evals_result)
eval_errors = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
# validation error should not decrease, if eta/learning_rate = 0
assert eval_errors[0] == eval_errors[-1]
# learning_rates as a customized decay function # learning_rates as a customized decay function
def eta_decay(ithround, num_boost_round): def eta_decay(ithround, num_boost_round):