def regression_metrics(self, predictions: DataFrame):
return ComputeModelStatistics(
evaluationMetric='regression',
labelCol=self.estimator.getLabelCol(),
scoresCol=self.estimator.getPredictionCol())\
.transform(predictions)\
.toPandas().to_dict(orient='list')
model = LightGBMRegressionModel.loadNativeModelFromFile("mymodel")
# COMMAND ----------
# MAGIC %md View the feature importances of the trained model.
# COMMAND ----------
print(model.getFeatureImportances())
# COMMAND ----------
# MAGIC %md Score the regressor on the test data.
# COMMAND ----------
scoredData = model.transform(test)
display(scoredData)
# COMMAND ----------
# MAGIC %md Compute metrics using ComputeModelStatistics
# COMMAND ----------
from mmlspark.train import ComputeModelStatistics
metrics = ComputeModelStatistics(evaluationMetric='regression',
labelCol='label',
scoresCol='prediction') \
.transform(scoredData)
display(metrics)
labelCol="income",
numFeatures=256).fit(train)
# COMMAND ----------
# MAGIC %md After the model is trained, we score it against the test dataset and view metrics.
# COMMAND ----------
from mmlspark.train import ComputeModelStatistics, TrainedClassifierModel
prediction = model.transform(test)
prediction.printSchema()
# COMMAND ----------
metrics = ComputeModelStatistics().transform(prediction)
metrics.limit(10).toPandas()
# COMMAND ----------
# MAGIC %md First, we will define the webservice input/output.
# MAGIC For more information, you can visit the [documentation for Spark Serving](https://github.com/Azure/mmlspark/blob/master/docs/mmlspark-serving.md)
# COMMAND ----------
from pyspark.sql.types import *
from mmlspark.io import *
import uuid
serving_inputs = spark.readStream.server() \
.address("localhost", 8898, "my_api") \
# COMMAND ----------
# MAGIC %md Score the regressor on the test data.
# COMMAND ----------
scoredData = model.transform(testCat)
scoredData.limit(10).toPandas()
# COMMAND ----------
# MAGIC %md Compute model metrics against the entire scored dataset
# COMMAND ----------
from mmlspark.train import ComputeModelStatistics
metrics = ComputeModelStatistics().transform(scoredData)
metrics.toPandas()
# COMMAND ----------
# MAGIC %md Finally, compute and show statistics on individual predictions in the test
# MAGIC dataset, demonstrating the usage of `ComputePerInstanceStatistics`
# COMMAND ----------
from mmlspark.train import ComputePerInstanceStatistics
evalPerInstance = ComputePerInstanceStatistics().transform(scoredData)
evalPerInstance.select("ArrDelay", "Scores", "L1_loss", "L2_loss") \
.limit(10).toPandas()
# MAGIC %md After the models have been trained and scored, compute some basic statistics
# MAGIC to evaluate the predictions. The following statistics are calculated for
# MAGIC regression models to evaluate:
# MAGIC * Mean squared error
# MAGIC * Root mean squared error
# MAGIC * R^2
# MAGIC * Mean absolute error
# MAGIC
# MAGIC Use the `ComputeModelStatistics` API to compute basic statistics for
# MAGIC the Poisson and the Random Forest models.
# COMMAND ----------
from mmlspark.train import ComputeModelStatistics
poissonMetrics = ComputeModelStatistics().transform(poissonPrediction)
print("Poisson Metrics")
poissonMetrics.toPandas()
# COMMAND ----------
randomForestMetrics = ComputeModelStatistics().transform(
randomForestPrediction)
print("Random Forest Metrics")
randomForestMetrics.toPandas()
# COMMAND ----------
# MAGIC %md We can also compute per instance statistics for `poissonPrediction`:
# COMMAND ----------
df_test = df_test.filter(col("label") != 2.0) \
.withColumn("label", when(col("label") > 0, 1.0).otherwise(0.0)) \
.select(["label", "text"])
print("Number of test samples after filtering: ", df_test.count())
# COMMAND ----------
# Make predictions
predictions = vw_trained.transform(df_test)
predictions.limit(10).toPandas()
# COMMAND ----------
# Compute model performance metrics
metrics = ComputeModelStatistics(evaluationMetric="classification",
labelCol="label",
scoredLabelsCol="prediction").transform(predictions)
metrics.toPandas()
# COMMAND ----------
# Utility class for plotting ROC curve (https://stackoverflow.com/questions/52847408/pyspark-extract-roc-curve)
class CurveMetrics(BinaryClassificationMetrics):
def __init__(self, *args):
super(CurveMetrics, self).__init__(*args)
def get_curve(self, method):
rdd = getattr(self._java_model, method)().toJavaRDD()
points = []
for row in rdd.collect():
points += [(float(row._1()), float(row._2()))]
# COMMAND ---------- # MAGIC %md Featurize images # COMMAND ---------- featurizedImages = cntkModel.transform(imagesWithLabels).select(["features","labels"]) # COMMAND ---------- # MAGIC %md Use featurized images to train a classifier # COMMAND ---------- from mmlspark.train import TrainClassifier from pyspark.ml.classification import RandomForestClassifier train,test = featurizedImages.randomSplit([0.75,0.25]) model = TrainClassifier(model=RandomForestClassifier(),labelCol="labels").fit(train) # COMMAND ---------- # MAGIC %md Evaluate the accuracy of the model # COMMAND ---------- from mmlspark.train import ComputeModelStatistics predictions = model.transform(test) metrics = ComputeModelStatistics(evaluationMetric="accuracy").transform(predictions) metrics.show()
import mmlspark
from mmlspark.lightgbm import LightGBMClassifier
from mmlspark.train import TrainClassifier, ComputeModelStatistics
df = spark.read.csv(r"C:\Users\yanrujing\Desktop\breast_cancer.csv",
header=True,
inferSchema=True)
train_data, test_data = df.randomSplit([0.8, 0.2], seed=0)
print(df.limit(10).toPandas())
model = TrainClassifier(model=LogisticRegression(),
labelCol="class",
numFeatures=256).fit(train_data)
prediction = model.transform(test_data)
metrics = ComputeModelStatistics().transform(prediction)
print(metrics.limit(10).toPandas())
f1 = VectorAssembler(inputCols=[
'clump_thickness', 'unif_cell_size', 'unif_cell_shape', 'marg_adhesion',
'single_epith_cell_size', 'bare_nuclei', 'bland_chrom', 'norm_nucleoli',
'mitoses'
],
outputCol='features')
f2 = StringIndexer(inputCol='class', outputCol='label')
p = Pipeline(stages=[f1, f2]).fit(df)
data = p.transform(df)
train_data, test_data = data.randomSplit([0.8, 0.2], seed=0)
model = LightGBMClassifier(objective='binary').fit(train_data)
test_predict = model.transform(test_data)
lr_predictions = lr_model.transform(lr_test_data)
display(lr_predictions.limit(10).toPandas())
# COMMAND ----------
# MAGIC %md We evaluate the prediction result by using `mmlspark.train.ComputeModelStatistics` which returns four metrics:
# MAGIC * [MSE (Mean Squared Error)](https://en.wikipedia.org/wiki/Mean_squared_error)
# MAGIC * [RMSE (Root Mean Squared Error)](https://en.wikipedia.org/wiki/Root-mean-square_deviation) = sqrt(MSE)
# MAGIC * [R quared](https://en.wikipedia.org/wiki/Coefficient_of_determination)
# MAGIC * [MAE (Mean Absolute Error)](https://en.wikipedia.org/wiki/Mean_absolute_error)
# COMMAND ----------
metrics = ComputeModelStatistics(
evaluationMetric='regression', labelCol='target',
scoresCol='prediction').transform(lr_predictions)
results = metrics.toPandas()
results.insert(0, 'model', ['Spark MLlib - Linear Regression'])
display(results)
# COMMAND ----------
# MAGIC %md ## Vowpal Wabbit
# COMMAND ----------
# MAGIC %md Perform VW-style feature hashing. Many types (numbers, string, bool, map of string to (number, string)) are supported.
# COMMAND ----------