experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Automated g4 rollback of changelist 166204326

Browse Source 
PiperOrigin-RevId: 166234847
This commit is contained in:
A. Unique TensorFlower 2017-08-23 12:05:25 -07:00 committed by TensorFlower Gardener
parent 0bfcea4b70
commit e2a4c81247
3 changed files with 30 additions and 59 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
tensorflow/contrib/boosted_trees/estimator_batch/trainer_hooks.py
View File

@ -114,7 +114,8 @@ class FeatureImportanceSummarySaver(session_run_hook.SessionRunHook):
summary_writer = SummaryWriterCache.get(output_dir)
usage_count_summary = Summary(value=[
Summary.Value(
tag="feature_importance/usage_counts", simple_value=usage_count)
tag="feature_importance/usage_counts",
simple_value=usage_count)
])
usage_fraction_summary = Summary(value=[
Summary.Value(
@ -123,14 +124,14 @@ class FeatureImportanceSummarySaver(session_run_hook.SessionRunHook):
])
summary_writer.add_summary(usage_count_summary, global_step)
summary_writer.add_summary(usage_fraction_summary, global_step)
gains_summary = Summary(value=[
Summary.Value(tag="feature_importance/gains", simple_value=gain)
])
gains_fraction_summary = Summary(value=[
Summary.Value(
gains_summary = Summary(
value=[Summary.Value(
tag="feature_importance/gains",
simple_value=gain)])
gains_fraction_summary = Summary(
value=[Summary.Value(
tag="feature_importance/gains_fraction",
simple_value=gain * gain_norm)
])
simple_value=gain * gain_norm)])
summary_writer.add_summary(gains_summary, global_step)
summary_writer.add_summary(gains_fraction_summary, global_step)
@ -143,7 +144,8 @@ class FeedFnHook(session_run_hook.SessionRunHook):
def before_run(self, run_context):
del run_context # unused by FeedFnHook.
return session_run_hook.SessionRunArgs(fetches=None, feed_dict=self.feed_fn)
return session_run_hook.SessionRunArgs(
fetches=None, feed_dict=self.feed_fn)
class StopAfterNTrees(session_run_hook.SessionRunHook):
@ -168,10 +170,8 @@ class StopAfterNTrees(session_run_hook.SessionRunHook):
num_finalized_trees = run_values.results["num_finalized_trees"]
assert num_attempted_trees is not None
assert num_finalized_trees is not None
# Stop when the required number of finalized trees is reached, or when we
# try enough times to build a tree but keep failing.
if (num_finalized_trees >= self._num_trees or
num_attempted_trees > 2 * self._num_trees):
num_attempted_trees > self._num_trees):
logging.info("Requesting stop since we have reached %d trees.",
num_finalized_trees)
run_context.request_stop()

14
tensorflow/contrib/boosted_trees/kernels/training_ops.cc
View File

@ -29,7 +29,6 @@ using boosted_trees::trees::TreeNode;
using boosted_trees::trees::TreeNodeMetadata;
using boosted_trees::utils::DropoutUtils;
using boosted_trees::learner::LearningRateConfig;
using boosted_trees::trees::Leaf;
namespace {
@ -45,11 +44,6 @@ struct SplitCandidate {
learner::SplitInfo split_info;
};
// Checks that the leaf is not empty.
bool IsLeafWellFormed(const Leaf& leaf) {
return leaf.has_sparse_vector() || leaf.has_vector();
}
// Helper method to update the best split per partition given
// a current candidate.
void UpdateBestSplit(
@ -71,14 +65,6 @@ void UpdateBestSplit(
return;
}
// If the current node is pure, one of the leafs will be empty, so the split
// is meaningless and we should not split.
if (!(IsLeafWellFormed(split->split_info.right_child()) &&
IsLeafWellFormed(split->split_info.left_child()))) {
VLOG(1) << "Split does not actually split anything";
return;
}
// Take the split if we don't have a candidate yet.
auto best_split_it = best_splits->find(partition_id);
if (best_split_it == best_splits->end()) {

51
tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch_test.py
View File

@ -617,8 +617,7 @@ class GbdtTest(test_util.TensorFlowTestCase):
learner_config.constraints.min_node_weight = 0
features = {}
batch_size = 3
features["dense_float"] = array_ops.constant(
[0.3, 1.5, 1.1], dtype=dtypes.float32)
features["dense_float"] = array_ops.ones([batch_size, 1], dtypes.float32)
gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel(
is_chief=True,
@ -687,25 +686,19 @@ class GbdtTest(test_util.TensorFlowTestCase):
stamp_token, serialized = model_ops.tree_ensemble_serialize(
ensemble_handle)
output.ParseFromString(serialized.eval())
self.assertEqual(len(output.trees), 1)
# We got 3 nodes: one parent and 2 leafs.
self.assertEqual(len(output.trees[0].nodes), 3)
self.assertEquals(len(output.trees), 1)
self.assertAllClose(output.tree_weights, [1])
self.assertEquals(stamp_token.eval(), 2)
# Leafs should have a dense vector of size 5.
expected_leaf_1 = [
-3.44797062874, -3.44291138649, 13.8490362167, -3.45003509521,
-3.45079040527
# Leaf should have a dense vector of size 5.
expected = [
-1.29972088337, -1.38769364357, 3.42812108994, 1.02690505981,
-1.75405228138
]
expected_leaf_2 = [
-1.25471234322, -1.31451582909, 1.52006089687, 2.38751673698,
-1.32643198967
]
self.assertAllClose(expected_leaf_1,
output.trees[0].nodes[1].leaf.vector.value)
self.assertAllClose(expected_leaf_2,
output.trees[0].nodes[2].leaf.vector.value)
for i in range(learner_config.num_classes):
self.assertAlmostEqual(expected[i],
output.trees[0].nodes[1].leaf.vector.value[i],
places=2)
def testTrainFnMulticlassDiagonalHessian(self):
"""Tests the GBDT train for multiclass diagonal hessian."""
@ -726,8 +719,7 @@ class GbdtTest(test_util.TensorFlowTestCase):
learner_config.constraints.min_node_weight = 0
batch_size = 3
features = {}
features["dense_float"] = array_ops.constant(
[0.3, 1.5, 1.1], dtype=dtypes.float32)
features["dense_float"] = array_ops.ones([batch_size, 1], dtypes.float32)
gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel(
is_chief=True,
@ -797,24 +789,17 @@ class GbdtTest(test_util.TensorFlowTestCase):
ensemble_handle)
output.ParseFromString(serialized.eval())
self.assertEqual(len(output.trees), 1)
# We got 3 nodes: one parent and 2 leafs.
self.assertEqual(len(output.trees[0].nodes), 3)
self.assertAllClose(output.tree_weights, [1])
self.assertEqual(stamp_token.eval(), 2)
# Leafs should have a dense vector of size 5.
expected_leaf_1 = [
-1.03539931774, -1.01067817211, 17.2976284027, -1.13134455681,
-4.50228404999
# Leaf should have a dense vector of size 5.
expected = [
-1.26767396927, -1.13043296337, 4.58542203903, 1.81428349018,
-2.43038392067
]
expected_leaf_2 = [
-1.29245102406, -1.13761842251, 2.20420837402, 3.10524249077,
-1.62693488598
]
self.assertAllClose(expected_leaf_1,
output.trees[0].nodes[1].leaf.vector.value)
self.assertAllClose(expected_leaf_2,
output.trees[0].nodes[2].leaf.vector.value)
for i in range(learner_config.num_classes):
self.assertAlmostEqual(expected[i],
output.trees[0].nodes[1].leaf.vector.value[i])
def testTrainFnMulticlassTreePerClass(self):
"""Tests the GBDT train for multiclass tree per class strategy."""