[Blocking] python kernel failed when call Booster().predict

[zzd1992]

Sometimes, the python kernel will be failed when I call Booster().predict method.
Here is an example:

import xgboost as xgb
import numpy as np

d = xgb.DMatrix(np.random.rand(1000, 5))
tree = xgb.Booster({'max_depth': 3, 'eta': 0.1, 'base_score': 0.0})

tree.boost(d, np.random.rand(1000), np.ones(1000))
print('start...')
preds = tree.predict(d)
print('end...')

Then

[19:35:05] src/tree/updater_prune.cc:74: tree pruning end, 1 roots, 14 extra nodes, 0 pruned nodes, max_depth=3
start...

Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)

However, when I modify the following line

tree.boost(d, np.random.rand(1000), np.ones(1000))

into

tree.boost(d, np.ones(1000), np.ones(1000))

Then every thing is ok.

start...
[19:49:16] src/tree/updater_prune.cc:74: tree pruning end, 1 roots, 0 extra nodes, 0 pruned nodes, max_depth=0
end...

How to fix this problem?

[trivialfis]

@zzd1992 Let me look into this.

[trivialfis]

@zzd1992 If you run the training process via xgb.train, XGBoost wiol throw an error tells you that you need labels for training since XGBoost implements supervised learning.

Maybe a better error handling in c++ is needed.

[zzd1992]

@trivialfis Thanks.
It doesn't work to add labels to DMatrix when I call the predict method.
And the predict method should not be dependent on labels.
In fact, I need to growth the tree one by one and modify its predictions by myself.
So it looks not suitable to run xgb.train. xgb.Booster().boost() is better for me.

[trivialfis]

@zzd1992 The label is only needed in training:

dtrain = xgb.DMatrix(train_data, label=labels)
bst = xgb.train(params, dtrain)

on prediction labels are not needed:

dtest = xgb.DMatrix(testing_data)
bst.predict(dtest)

The error you encountered is due to the fact that with empty labels in DMatrix, no training can be performed hence no tree is built. Later prediction needs at least one tree, which leads to an memory error.

[zzd1992]

@trivialfis After running Booster().boost(), one tree is built. So there already exists a tree when predicting.
Here is an example:

d = xgb.DMatrix(np.random.rand(1000, 5), np.random.rand(1000))
tree = xgb.Booster({'max_depth': 3, 'eta': 0.1, 'base_score': 0.0})

tree.boost(d, np.random.rand(1000), np.ones(1000))
print(tree.get_dump()[0])

print('start...')
preds = tree.predict(d)
print('end...')

Then

0:[f3<0.011232] yes=1,no=2,missing=1
	1:[f0<0.817435] yes=3,no=4,missing=3
		3:leaf=-0.0150572
		4:leaf=-0.00429241
	2:[f4<0.30253] yes=5,no=6,missing=5
		5:[f4<0.291276] yes=7,no=8,missing=7
			7:leaf=-0.048414
			8:leaf=-0.0217552
		6:[f4<0.997218] yes=9,no=10,missing=9
			9:leaf=-0.0531961
			10:leaf=-0.0129695

start...
[23:24:37] src/tree/updater_prune.cc:74: tree pruning end, 1 roots, 10 extra nodes, 0 pruned nodes, max_depth=3

Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)

[trivialfis]

I never tried this before. I don't think Booster.boost or Booster.update is meant to be called by users. But still segmentation fault is not acceptable.