H2O GAM - 加权：预测不再有效

Question

如果我训练加权 H2O GAM 回归模型，我无法用它进行预测。使用参数 weights_column

完成加权回归

我是运行python=3.6.13，h2o=3.32.1.3，pandas=0.25.3，numpy=1.19.5，sklearn=0.24.2。 Java版本：openjdk版本“14.0.2”。

预测适用于：

未加权的 H2O GAM
加权 H2O GLM
降级到 h2o 时加权 H2O GAM=3.32.0.5

我已将此注册为 http://jira.h2o.ai 上的错误，但如果有人有办法让它工作而不降级 h2o，我仍然会感兴趣。

import numpy as np
import pandas as pd
from sklearn.datasets import load_boston
import h2o
from h2o.estimators.gam import H2OGeneralizedAdditiveEstimator

h2o.no_progress()
h2o.init()

np.random.seed(42)
boston = load_boston()
y = pd.Series(boston["target"], name="y")
X = pd.DataFrame(boston["data"], columns=boston["feature_names"])  # shape: (506, 13)
myweight = pd.Series(np.random.random_sample((len(y),)), name="myweight2")

predictors = ['CRIM', 'AGE']
gam_columns = ['CRIM']

params = {
    "family": "gaussian",
    "gam_columns": gam_columns,
    'bs': len(gam_columns) * [0],
}

df0 = pd.concat([y, X, myweight], axis=1)
df = h2o.H2OFrame(python_obj=df0)

model = H2OGeneralizedAdditiveEstimator(**params)
model.train(
    x=predictors,
    y="y",
    weights_column="myweight2",
    training_frame=df,
)

print('df.shape', df.shape)
y_pred = model.predict(df)
print('y_pred:', y_pred.as_data_frame()["predict"].values[0:5])

我得到了这个输出。它抱怨 myweight2:

Checking whether there is an H2O instance running at http://localhost:54321 . connected.
--------------------------  ------------------------------------------

df.shape (506, 15)
Traceback (most recent call last):
  File "/Users/g009655/tmp7/h2otest/test_gam_predict.py", line 37, in <module>
    y_pred = model.predict(df)
  File "/Users/g009655/Library/Caches/pypoetry/virtualenvs/h2otest-S7Xak4Mg-py3.6/lib/python3.6/site-packages/h2o/model/model_base.py", line 237, in predict
    j.poll()
  File "/Users/g009655/Library/Caches/pypoetry/virtualenvs/h2otest-S7Xak4Mg-py3.6/lib/python3.6/site-packages/h2o/job.py", line 80, in poll
    "\n{}".format(self.job_key, self.exception, self.job["stacktrace"]))
OSError: Job with key 017f00000132d4ffffffff$_9242dd1b28497090cf9ccad52bd54b9f failed with an exception: java.lang.AssertionError:  null vec: ff0f000000ffffffff$_b0f0839f8f1a041e8bf5254b552e4dd3; 

name: myweight2

stacktrace: 
java.lang.AssertionError:  null vec: ff0f000000ffffffff$_b0f0839f8f1a041e8bf5254b552e4dd3; 

name: myweight2

    at water.fvec.Frame.<init>(Frame.java:161)
    at hex.gam.GAMModel.cleanUpInputFrame(GAMModel.java:505)
    at hex.gam.GAMModel.adaptTestForTrain(GAMModel.java:492)
    at hex.Model.score(Model.java:1697)
    at water.api.ModelMetricsHandler.compute2(ModelMetricsHandler.java:422)
    at water.H2O$H2OCountedCompleter.compute(H2O.java:1637)
    at jsr166y.CountedCompleter.exec(CountedCompleter.java:468)
    at jsr166y.ForkJoinTask.doExec(ForkJoinTask.java:263)
    at jsr166y.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:974)
    at jsr166y.ForkJoinPool.runWorker(ForkJoinPool.java:1477)
    at jsr166y.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:104)

Closing connection _sid_ad95 at exit
H2O session _sid_ad95 closed.

Process finished with exit code 1

Answer 1

感谢您的错误报告。这里有一张link的Jira票据，供参考。

Answer 2

我遇到同样的错误，但我找到了解决方法。对于我重新加载（在我的情况下来自 pandas.DataFrame），培训 H2OFrame 有效。似乎在训练中它以某种方式损坏了...

对于你的情况，尝试：

df = h2o.H2OFrame(python_obj=df0)
y_pred = model.predict(df)

H2O GAM - 加权：预测不再有效

H2O GAM - weighted: prediction does not work anymore

python

predict

gam

h2o