XGBoost Spark One Model Per Worker Integration

Question

Trying to work through this notebook https://databricks-prod-cloudfront.cloud.databricks.com/public/4027ec902e239c93eaaa8714f173bcfc/1526931011080774/3624187670661048/6320440561800420/latest.html.

Using spark version 2.4.3 and xgboost 0.90

Keep getting this error ValueError: bad input shape () when trying to execute ...

features = inputTrainingDF.select("features").collect()
lables = inputTrainingDF.select("label").collect()

X = np.asarray(map(lambda v: v[0].toArray(), features))
Y = np.asarray(map(lambda v: v[0], lables))

xgbClassifier = xgb.XGBClassifier(max_depth=3, seed=18238, objective='binary:logistic')

model = xgbClassifier.fit(X, Y)
ValueError: bad input shape ()

and

def trainXGbModel(partitionKey, labelAndFeatures):
  X = np.asarray(map(lambda v: v[1].toArray(), labelAndFeatures))
  Y = np.asarray(map(lambda v: v[0], labelAndFeatures))
  xgbClassifier = xgb.XGBClassifier(max_depth=3, seed=18238, objective='binary:logistic' )
  model =  xgbClassifier.fit(X, Y)
  return [partitionKey, model]

xgbModels = inputTrainingDF\
.select("education", "label", "features")\
.rdd\
.map(lambda row: [row[0], [row[1], row[2]]])\
.groupByKey()\
.map(lambda v: trainXGbModel(v[0], list(v[1])))

xgbModels.take(1)
ValueError: bad input shape ()

You can see in the notebook it is working for whoever posted it. My guess is it has something to do with the X and Y np.asarray() mapping because the logic is just trying to map the label and features to the function but the shapes are empty. Got it working using this code

pandasDF = inputTrainingDF.toPandas()
series = pandasDF['features'].apply(lambda x : np.array(x.toArray())).as_matrix().reshape(-1,1)
features = np.apply_along_axis(lambda x : x[0], 1, series)
target = pandasDF['label'].values
xgbClassifier = xgb.XGBClassifier(max_depth=3, seed=18238, objective='binary:logistic' )
model = xgbClassifier.fit(features, target)

however want to integrate into the original function call & understand why the original notebook does not work. An extra set of eyes to troubleshoot this would be much appreciated!

prudenko prudenko · Accepted Answer · 2019-11-05T10:05:16

You are probably using python3. The issue is that in python3 map function returns an iterator object, rather than a collection. All you have to do to fix this example is to change map -> list(map(...)):

def trainXGbModel(partitionKey, labelAndFeatures):
  X = np.asarray(list(map(lambda v: v[1].toArray(), labelAndFeatures)))
  Y = np.asarray(list(map(lambda v: v[0], labelAndFeatures)))

Or you can use np.fromiter to convert iterable object to numpy array.

XGBoost Spark One Model Per Worker Integration

1 Answers