def gridsearch(XX, XXpredict, yy, yypredict, clf):
# tuned_parameters=settings.param_grid
param_grid = settings.param_grid
print("Gridsearch start")
def report(grid_scores, n_top=3):
top_scores = sorted(grid_scores, key=itemgetter(1),
reverse=True)[:n_top]
for i, score in enumerate(top_scores):
print("Model with rank: {0}".format(i + 1))
print("Mean validation score: {0:.3f} (std: {1:.3f})".format(
score.mean_validation_score,
numpy.std(score.cv_validation_scores)))
print("Parameters: {0}".format(score.parameters))
print("")
grid_search = GridSearchCV(sc,
clf,
param_grid=param_grid,
cv=10,
n_jobs=-1,
verbose=1)
start = time()
grid_search.fit(XX, yy)
print(
"GridSearchCV took {:.2f} seconds for {:d} candidate settings.".format(
time() - start, len(grid_search.grid_scores_)))
report(grid_search.grid_scores_)
return grid_search
def grid_search_svm(X_train, y_train,X_test,ngrams,n_split,svm_choice='linear',tfidf_choice=False,nums_train=None,nums_test=None):
svm=None
grid=None
if svm_choice == 'linear':
svm = LinearSVC()
c_array = np.logspace(1., 4., num=4)
if tfidf_choice:
grid = {'vect__ngram_range': ngrams, 'tfidf__use_idf': (True, False),
'clf__C': c_array.tolist()}
else:
grid = {'vect__ngram_range': ngrams,
'clf__C': c_array.tolist()}
elif svm_choice == 'svc':
svm = SVC()
c_array = np.logspace(-3., 6., num=10)
g_array = np.logspace(-3., 3., num=7)
if tfidf_choice:
grid = {'vect__ngram_range': ngrams,
'tfidf__use_idf': (True, False),
'clf__kernel': ['rbf'],
'clf__C': c_array.tolist(),
'clf__gamma': g_array.tolist()}
else:
grid = {'vect__ngram_range': ngrams,
'clf__kernel': ['rbf'],
'clf__C': c_array.tolist(),
'clf__gamma': g_array.tolist()}
if type(nums_train) is np.ndarray and type(nums_test) is np.ndarray:
if tfidf_choice:
clf_pipeline = Pipeline([('vect', CountVectorizer(ngram_range=ngrams)),
('tfidf', TfidfTransformer(smooth_idf=False)),
('numfeat', NumFeatureAdder(nums_train,nums_test)),
('clf',svm)])
else:
clf_pipeline = Pipeline([('vect', CountVectorizer(ngram_range=ngrams)),
('numfeat', NumFeatureAdder(nums_train, nums_test)),
('clf', svm)])
else:
if tfidf_choice:
clf_pipeline = Pipeline([('vect', CountVectorizer(ngram_range=ngrams)),
('tfidf', TfidfTransformer(smooth_idf=False)),
('clf',svm)])
else:
clf_pipeline = Pipeline([('vect', CountVectorizer(ngram_range=ngrams)),
('clf',svm)])
print(clf_pipeline.get_params().keys())
sc = SparkContext.getOrCreate()
grid_search = GridSearchCV(sc, clf_pipeline, grid, n_jobs=-1, cv=n_split)
grid_search.fit(X_train, y_train)
grid_search_predicted = grid_search.predict(X_test)
return grid_search_predicted
def grid_search(sc, data, label, features):
"""
使用grid search寻找最优的超参数
"""
# 产生备选的超参数集
parameters = {"alpha": 10**np.linspace(-4, 0, 45)}
# Lasso模型里有超参数alpha，表示惩罚项的权重
la = Lasso()
gs = GridSearchCV(sc, la, parameters)
gs.fit(data[features], data[label])
return gs
def gridSearch(sc, data, label, features):
"""
使用 grid search 寻找最优的超参数
:param sc:
:param data:
:param label:
:param features:
:return:
"""
parameters = {"alpha": 10**np.linspace(-4, 0, 45)}
la = Lasso()
gs = GridSearchCV(sc, la, parameters)
gs.fit(data[features], data[label])
return gs
def grid_search(sc, X, Y, K):
roc_auc_scorer = make_scorer(roc_auc_score)
clf = SPGridSearchCV(sc,
pipe,
cv=K,
n_jobs=2,
param_grid=param_grid,
scoring=roc_auc_scorer)
clf.fit(X, Y)
grid_result = clf.grid_scores_
param_score = []
for i in range(len(grid_result)):
param_score.append((grid_result[i][0], grid_result[i][1]))
print(param_score)
return param_score
class Classifier:
def __init__(self):
self.model = None
def preprocess(self, data):
''' Optional for preprocessing'''
return CountVectorizer().fit_transform(data)
def train(self,X, y, method="rf"):
param_grid = {
"max_depth": [6, None],
"max_features": [5,10,20],
}
obj = RandomForestClassifier()
if method == "svm":
obj = SVC()
self.model = GridSearchCV(RandomForestClassifier(), param_grid=param_grid)
self.model.fit(X, y)
def preidict(self, X):
if self.model is None:
return
return self.model.predict(X)
class Classifier:
def __init__(self):
self.model = None
def preprocess(self, data):
''' Optional for preprocessing'''
return CountVectorizer().fit_transform(data)
def train(self, X, y, method="rf"):
param_grid = {
"max_depth": [6, None],
"max_features": [5, 10, 20],
}
obj = RandomForestClassifier()
if method == "svm":
obj = SVC()
self.model = GridSearchCV(RandomForestClassifier(),
param_grid=param_grid)
self.model.fit(X, y)
def preidict(self, X):
if self.model is None:
return
return self.model.predict(X)
def main():
"""
main function, runs the program
trains spark sklearn model
"""
absolute_path = "/data/model_data/"
train_df = np.loadtxt(absolute_path + "train.csv", delimiter=',')
train_target_df = np.loadtxt(absolute_path + "target_train.csv", delimiter=',')
test_df = np.loadtxt(absolute_path + "test.csv", delimiter=',')
test_target_df = np.loadtxt(absolute_path + "target_test.csv", delimiter=',')
regr = RandomForestRegressor(random_state=0, n_estimators=1000, min_samples_leaf=1) # best model so far!
# pyspark
regr_rf_cv = GridSearchCV(sc=spark.sparkContext,
estimator=regr,
n_jobs=20,
cv=5,
verbose=5,
param_grid={})
regr_rf_cv.fit(train_df, train_target_df)
y_list, y_hat_list = run_test(test_df, test_target_df, regr_rf_cv)
print("Mean absolute error: {}".format(get_mean_absolute_error(y_list, y_hat_list)))
print("Average relative error: {}".format(get_average_relative_error(y_list, y_hat_list)))
save_model(regr_rf_cv.best_estimator_, "rf_uber_model", "/data/saved_model/")
load_model("/data/saved_model/rf_uber_model.pkl", testExample=(test_df[0], test_target_df[0]))
def with_spark(x_train, x_test, y_train, y_test, vents_and_files, spark_connect_str):
from pyspark import SparkConf, SparkContext
from spark_sklearn import GridSearchCV as SparkGridSearchCV
conf = SparkConf().setMaster(spark_connect_str).setAppName("ecs251")
conf.set("spark.executor.memory", "1g")
conf.set("spark.serializer", "org.apache.spark.serializer.KryoSerializer")
conf.set("spark.deploy.mode", "cluster")
sc = SparkContext(conf=conf)
param_grid = {"C": C, "gamma": GAMMA}
gs = SparkGridSearchCV(sc, SVC(cache_size=CACHE_SIZE), param_grid=param_grid)
res = gs.fit(x_train, y_train['y'].values)
print("Best Score: ", res.best_score_)
print("Best params: ", res.best_params_)
predictions = res.predict(x_test)
print("Accuracy: " + str(accuracy_score(y_test['y'], predictions)))
print("Precision: " + str(precision_score(y_test['y'], predictions)))
print("Recall: " + str(recall_score(y_test['y'], predictions)))
fpr, tpr, thresh = roc_curve(y_test['y'], predictions)
print("False pos rate: " + str(fpr[1]))
print("True post rate: " + str(tpr[1]))
def test():
import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.ensemble import GradientBoostingClassifier
# from sklearn.model_selection import GridSearchCV
from spark_sklearn import GridSearchCV
from pyspark import SparkConf, SparkContext, HiveContext
from spark_sklearn import Converter
import time
start = time.time()
conf = SparkConf().setAppName("spark-sklearn")
sc = SparkContext(conf=conf)
spark = HiveContext(sc)
path = "/home/data/data_cell_lable_0521_rsrp_five3_all.csv"
df = spark.read.csv(path, header=True, inferSchema=True)
converter = Converter(sc)
df_data = converter.toPandas(df)
# 也可以直接使用 pandas的DataFrame进行操作
# inputpath1 = '/home/etluser/xiexiaoxuan/data/data_cell_lable_0521_rsrp_five3_all.csv'
# df_data = pd.read_csv(inputpath1)
df_data = df_data.dropna(axis=0, how='any')
x1 = df_data.drop(['label'], axis=1)
y1 = df_data['label']
gbm0 = GradientBoostingClassifier(n_estimators=262,
max_depth=57,
min_samples_split=50,
random_state=10,
subsample=0.7,
learning_rate=0.01)
pipeline = Pipeline([("standard", StandardScaler()), ("gbdt", gbm0)])
params = {
"gbdt__n_estimators": [i for i in range(10, 20)],
"gbdt__max_depth": [i for i in range(3, 20)]
}
grid_search = GridSearchCV(sc,
pipeline,
param_grid=params,
error_score=0,
scoring="accuracy",
cv=5,
n_jobs=10,
pre_dispatch="2*n_jobs",
return_train_score=False)
grid_search.fit(x1, y1)
end = time.time()
print("总耗时 ：%.2f s" % (end - start))
print(grid_search.best_estimator_)
index = grid_search.best_index_
res = grid_search.cv_results_
best_score = res["mean_test_score"][index]
print("===============: " + str(best_score))
# Utility function to report best scores
def report(grid_scores, n_top=3):
top_scores = sorted(grid_scores, key=itemgetter(1), reverse=True)[:n_top]
for i, score in enumerate(top_scores):
print("Model with rank: {0}".format(i + 1))
print("Mean validation score: {0:.3f} (std: {1:.3f})".format(
score.mean_validation_score,
np.std(score.cv_validation_scores)))
print("Parameters: {0}".format(score.parameters))
print("")
# create spark context - use all cores
sc = SparkContext("local[*]", "Simple Test")
digits = datasets.load_digits()
X, y = digits.data, digits.target
param_grid = {"max_depth": [3, 5, 8, 10, 13, None],
"max_features": [1, 3, 10, 20],
"min_samples_split": [1, 3, 10],
"min_samples_leaf": [1, 3, 10],
"criterion": ["gini", "entropy"],
"n_estimators": [500]}
clf = RandomForestClassifier()
#gs = grid_search.GridSearchCV(clf, param_grid=param_grid) # local
gs = GridSearchCV(sc, clf, param_grid=param_grid) # distributed
start = time()
gs.fit(X, y)
print("GridSearchCV took {:.2f} seconds for {:d} candidate settings.".format(time() - start, len(gs.grid_scores_)))
report(gs.grid_scores_)
from dl_steer import dt_handler, coordinator, custom_model, engine_interface, provenance
from keras.wrappers.scikit_learn import KerasClassifier
from spark_sklearn import GridSearchCV
from keras.models import Sequential
data = dt_handler.read_dataset('input_data.csv')
...
model = KerasClassifier(build_fn=custom_model.get_model(), verbose=0)
X, y = data['X'], data['y']
queue = coordinator.get_queue()
for hyperparameter_combination in queue:
provenance.persist(hyperparameter_combination)
grid = GridSearchCV(estimator=model,
param_grid=hyperparameter_combination,
n_jobs=-1,
scoring="accuracy")
grid_result = grid.fit(X, y)
provenance.persist(grid_result)
#The method below verifies if user steered the queue. If yes, it reloads the queue accordingly.
queue.checkSteering()
sc = SparkContext(conf=conf)
digits = load_digits()
n_samples = len(digits.images)
data = digits.images.reshape((n_samples, -1))
X_train, X_test, y_train, y_test = train_test_split(data, digits.target, test_size=0.3, random_state=0)
svc = svm.SVC()
hyperparam_grid = {
'kernel': ['linear', 'poly', 'rbf', 'sigmoid'],
'gamma': np.linspace(0.001, 0.01, num=10),
'C': np.linspace(1, 10, num=10),
'tol': np.linspace(0.01, 0.1, 10)
}
classifier = GridSearchCV(sc, svc, hyperparam_grid)
start = time()
classifier.fit(X_train, y_train)
elapsed = time() - start
print('elapsed: {} seconds'.format(elapsed))
print('Best Kernel:\t{}'.format(classifier.best_estimator_.kernel))
print('Best Gamma:\t{}'.format(classifier.best_estimator_.gamma))
print('Best C:\t\t{}'.format(classifier.best_estimator_.C))
y_pred = classifier.predict(X_test)
print('Accuracy:\t{:.1%}'.format(metrics.accuracy_score(y_test, y_pred)))
# Create hold-out test dataset
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.25)
param_grid = {
"max_depth": [3, None],
"max_features": [1, 3, 10],
"min_samples_leaf": [1, 3, 10],
"bootstrap": [True, False],
"criterion": ["gini", "entropy"],
"n_estimators": [10, 20, 40, 80]
}
gs = GridSearchCV(sc=sc,
estimator=RandomForestClassifier(),
cv=4,
param_grid=param_grid,
refit=True)
with timeit():
gs.fit(x_train, y_train)
results = pd.DataFrame(gs.cv_results_)
print(results.sort_values(['mean_test_score'], ascending=False)[0:10])
# Validate accuracy of best model against hold-out data
best_model = gs.best_estimator_
test_accuracy = best_model.score(x_test, y_test)
print(test_accuracy)
logger.log('Best model accuracy', test_accuracy)
from spark_sklearn import GridSearchCV
from sklearn.model_selection import TimeSeriesSplit
from sklearn.metrics import make_scorer, mean_absolute_error, r2_score
lasso_run = \
GridSearchCV(sc,
estimator=get_lasso_pipeline(),
param_grid={'lso__normalize':[True,False],
'lso__alpha' :[10.0**n for n in range(-3,4)]},
cv=TimeSeriesSplit(n_splits=5),
scoring=make_scorer(r2_score),
return_train_score=False,
n_jobs=-1
)
lasso_run.fit(trn_coal_cnt_fea_pdf, trn_coal_cnt_tgt_ser)
display_pdf(est_grid_results_pdf(lasso_run,
est_tag='lasso'))
# COMMAND ----------
lasso_run.fit(trn_ore_tfidf_fea_pdf, trn_ore_tfidf_tgt_ser)
display_pdf(est_grid_results_pdf(lasso_run,
est_tag='lasso'))
# COMMAND ----------
# MAGIC %md The scores for the model with the best hyperparameters for the model trained and validated with the two datasets are:<table>
# MAGIC <tr>
######################################################
# Training
#######################################################
tuned_parameters = [{'n_estimators': [300,400], 'max_depth': [3,4], 'objective': ['multi:softprob'],
'reg_alpha': [1], 'reg_lambda': [1], 'colsample_bytree': [1],
'learning_rate': [0.1], 'colsample_bylevel': [0.01,0.1],
'subsample': [0.5,0.7]}]
clf = XGBClassifier(seed=0)
metric = 'neg_log_loss'
sc = SparkContext.getOrCreate()
clf_cv = GridSearchCV(sc = sc, param_grid = tuned_parameters, estimator = clf,
scoring=metric, cv=5, verbose=3)
model = clf_cv.fit(X_train,y_train)
run_logger.log(metric, float(clf_cv.best_score_))
for key in clf_cv.best_params_.keys():
run_logger.log(key, clf_cv.best_params_[key])
if not path.exists('./outputs'):
makedirs('./outputs')
outfile = open('./outputs/sweeping_results.txt','w')
print("metric = ", metric, file=outfile)
for i in range(len(model.grid_scores_)):
print(model.grid_scores_[i], file=outfile)
outfile.close()
# In[99]:
tuned_parameters = {
"n_estimators": [ 100 ],
"max_depth" : [ 3 ],
"learning_rate": [ 0.1 ],
}
gbc = ensemble.GradientBoostingClassifier()
clf = GridSearchCV(spark.sparkContext, gbc, tuned_parameters)
clf
# In[100]:
clf.fit(X_timetrain, Y_timetrain_arr)
clftest_pred = clf.predict(X_timetest)
print "Accuracy is ", metrics.accuracy_score(Y_timetest_arr, clftest_pred) *100, "%"
# In[101]:
knn1 = KNeighborsClassifier()
knn_params = {
"n_neighbors": [31]
}
clf2 = GridSearchCV(spark.sparkContext, knn1, knn_params, n_jobs = 2)
clf2
# In[102]:
from sklearn import grid_search
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from spark_sklearn import GridSearchCV
df = pd.read_csv('../data/master_RushPassOnly.csv')
y = df.pop('IsPass').values
X = df.values
param_grid = {
"max_depth": [3, 5, 10, None],
"max_features": [None, 'auto', 'log2'],
"n_estimators": [100],
'min_samples_split': [2, 4],
'min_samples_leaf': [1, 2, 4],
'bootstrap': [True, False]
}
rf = RandomForestClassifier(verbose=2, n_jobs=-1)
gs = GridSearchCV(rf,
param_grid=param_grid,
n_jobs=-1,
verbose=2,
scoring='neg_mean_squared_error')
gs.fit(X, y)
best_parameters = gs.best_params_
print best_parameters
SPARK_HOME + 'python/lib/pyspark.zip',
SPARK_HOME + 'python/lib/py4j-0.10.1-src.zip']
)
from pyspark import SparkContext
from pyspark import SparkConf
if __name__ == '__main__':
conf = SparkConf()
conf.setMaster("local[3]")
# 指定具体的Master机器 地址和端口
# conf.setMaster("spark://jdwang-HP:7077")
conf.setAppName("spark_test")
# 可以设置属性等
# conf.set("spark.executor.memory", "12g")
sc = SparkContext(conf=conf)
# 测试
from sklearn import svm, datasets
from spark_sklearn import GridSearchCV
iris = datasets.load_iris()
parameters = {'kernel': ('linear', 'rbf'), 'C': [1, 10]}
svr = svm.SVC()
clf = GridSearchCV(sc, svr, parameters)
clf.fit(iris.data, iris.target)
print(clf.best_params_)
print(clf.predict(iris.data))
end_time = time.time()
print('running time is %ds'%(end_time-start_time))
# COMMAND ----------
# MAGIC %md Create SVC Model
# COMMAND ----------
from sklearn import svm, grid_search, datasets
from spark_sklearn import GridSearchCV
parameters = {
'kernel': ('linear', 'rbf', 'poly', 'rbf', 'sigmoid'),
'C': [1, 20]
}
svr = svm.SVC()
clf = GridSearchCV(sc, svr, param_grid=parameters, scoring='accuracy')
clf.fit(x_train, y_train)
print(clf.best_params_)
bestsvc = clf.best_estimator_
print(clf.best_score_)
# COMMAND ----------
# MAGIC %md Create Random Forest Model
# COMMAND ----------
en_rf = RandomForestClassifier(n_estimators=64,
max_depth=32,
min_samples_split=128,
random_state=0)
X_minus_trea = X[np.where(y != 'TREA')]
y_minus_trea = y[np.where(y != 'TREA')]
X_final = X_minus_trea[np.where(y_minus_trea != 'PORNOGRAPHY/OBSCENE MAT')]
y_final = y_minus_trea[np.where(y_minus_trea != 'PORNOGRAPHY/OBSCENE MAT')]
# Separate training, dev, and test data:
test_data, test_labels = X_final[800000:], y_final[800000:]
dev_data, dev_labels = X_final[700000:800000], y_final[700000:800000]
train_data, train_labels = X_final[100000:700000], y_final[100000:700000]
calibrate_data, calibrate_labels = X_final[:100000], y_final[:100000]
# Create mini versions of the above sets
mini_train_data, mini_train_labels = X_final[:20000], y_final[:20000]
mini_calibrate_data, mini_calibrate_labels = X_final[19000:28000], y_final[
19000:28000]
mini_dev_data, mini_dev_labels = X_final[49000:60000], y_final[49000:60000]
param_grid = {
'C': [.001, .01, .01] + [i for i in range(1, 100, 5)],
"penalty": ['l1', 'l2']
}
clf = LogisticRegression()
gs = GridSearchCV(sc, clf, param_grid)
start = time()
gs.fit(mini_train_data, mini_train_labels)
print("GridSearchCV took {:.2f} seconds for {:d} candidate settings.".format(
time() - start, len(gs.grid_scores_)))
report(gs.grid_scores_)
from sklearn.ensemble import RandomForestRegressor
from sklearn.datasets import load_boston
from spark_sklearn import GridSearchCV
import pyspark
if __name__ == '__main__':
sc = pyspark.SparkContext('local[*]')
boston = load_boston()
RAMDON_FOREST_PARAMS = {
"n_estimators": [100],
"max_features": [1, "auto", "sqrt", None],
"max_depth": [1, 5, 10, None],
"min_samples_leaf": [1, 2, 4, 50]
}
rf = RandomForestRegressor(random_state=0, n_jobs=-1)
clf = GridSearchCV(sc, rf, RAMDON_FOREST_PARAMS)
clf.fit(boston.data, boston.target)
print("parameters for random forest: {0}".format(clf.best_params_),
sep="\n")
documentDF = session.createDataFrame([
("Hi I heard about Spark", "spark"),
("I wish Java could use case classes", "java"),
("Logistic regression models are neat", "mlib"),
("Logistic regression models are neat", "spark"),
("Logistic regression models are neat", "mlib"),
("Logistic regression models are neat", "java"),
("Logistic regression models are neat", "spark"),
("Logistic regression models are neat", "java"),
("Logistic regression models are neat", "mlib")
], ["text", "preds"]).select(f.split("text", "\\s+").alias("new_text"), "preds")
word2vec = Word2Vec(vectorSize=100, minCount=1, inputCol="new_text",
outputCol="features")
indexer = StringIndexer(inputCol="preds", outputCol="labels")
pipline = Pipeline(stages=[word2vec, indexer])
ds = pipline.fit(documentDF).transform(documentDF)
data = ds.toPandas()
parameters = {'kernel': ('linear', 'rbf')}
svr = svm.SVC()
clf = GridSearchCV(session.sparkContext, svr, parameters)
X = [x.values for x in data.features.values]
y = [int(x) for x in data.labels.values]
model = clf.fit(X, y)
# modelB = session.sparkContext.broadcast(pickle.dumps(model))
# wow = documentDF.rdd.map(lambda row: pickle.loads(modelB.value).transform(row["features"].values)).collect()
# print(wow)
from pyspark import SparkContext
from sklearn import svm, datasets
from spark_sklearn import GridSearchCV
# create spark context - use all cores
sc = SparkContext("local[*]", "Simple Test")
# test data
iris = datasets.load_iris()
parameters = {'kernel':('linear', 'rbf'), 'C':[0.01, 0.1, 1, 2, 3, 4, 5, 10]}
svr = svm.SVC()
clf = GridSearchCV(sc, svr, parameters)
clf.fit(iris.data, iris.target)
print(clf.best_estimator_)
print(clf.best_score_)
'max_depth': 6,
'subsample': 0.7,
'colsample_bytree': 0.7,
'objective': 'reg:linear',
'silent': 1
}
import xgboost as xgb
dtrain = xgb.DMatrix(train_X, train_y, feature_names=train_X.columns.values)
model = xgb.train(dict(xgb_params, silent=0), dtrain, num_boost_round=100, feval=xgb_r2_score, maximize=True)
# Gradient Boosting Regressor
gbr = ensemble.GradientBoostingRegressor()
clf = GridSearchCV(gbr, cv=3, param_grid=tuned_parameters,
scoring='median_absolute_error')
preds = clf.fit(X_train, y_train)
best = clf.best_estimator_
# plot error for each round of boosting
# Note: best_estimator_, staged_predict
test_score = np.zeros(n_est, dtype=np.float64)
train_score = best.train_score_
for i, y_pred in enumerate(best.staged_predict(X_test)):
test_score[i] = best.loss_(y_test, y_pred)
### Grid search
from pyspark import SparkContext, SparkConf
from spark_sklearn import GridSearchCV
conf = SparkConf()
def generate_model_package(training_data_path, id_cols, target_cols,
fields_config_file, param_grid, model_name,
target_var):
"""
training_data_path
,id_cols
,target_cols
,fields_config_file
,param_grid
,model_name
,target_var
"""
pyspark_app_nm = "train_" + model_name + "_" + secrets.token_hex(nbytes=4)
logging.info("Starting process: " + pyspark_app_nm)
#create spark object and spark context for parallel learning
logging.info("Instantiating pyspark.")
app_pyspark_conf = SparkConf()
app_pyspark_conf.setAppName(pyspark_app_nm)
# app_pyspark_conf.set('spark.executor.memory',spark_executor_memory)
# app_pyspark_conf.set('spark.executor.cores', spark_executor_cores)
spark = SparkSession.builder.config(conf=app_pyspark_conf).getOrCreate()
sc = spark.sparkContext
#load data
logging.info("Beginning data load.")
training_df = pd.read_parquet(training_data_path, engine='pyarrow')
# sampling down
# training_df_1 = training_df[training_df[target_var]==1].sample(20)
# training_df_0 = training_df[training_df[target_var]==0].sample(40)
# training_df = pd.concat([training_df_0,training_df_1])
# column handling
logging.info("Creating column lists")
all_cols = training_df.columns.tolist()
x_cols = list(set(all_cols) - (set(target_cols + id_cols)))
# dataframe setup
X = training_df[x_cols]
y = training_df[target_cols]
# create holdout data
logging.info("Creating holdout data")
x_train, x_test, y_train, y_test = train_test_split(X,
y[target_var],
test_size=0.1,
stratify=y[target_var])
wts = y_test.value_counts()
wtrat = (wts[0] / wts[1])
# instantiate model
gbm = lgb.LGBMClassifier()
fit_params = {
"eval_set": [(x_test, y_test)],
"eval_metric": ear_stop_eval_mtr,
"early_stopping_rounds": ear_stop_rnds
# ,"scale_pos_weight": wtrat
}
grid_search = SparkGridSearchCV(sc,
estimator=gbm,
param_grid=param_grid,
fit_params=fit_params)
# grid_search.fit(x_train,y_train)
grid_search.fit(x_train, y_train)
best_model = grid_search.best_estimator_
optimized_parameters = best_model.get_params()
# create confusion dataframe
y_true = pd.DataFrame(y_test)
y_true = y_true.reset_index()
y_true.columns.values[0] = "CUSTOMER_KEY"
y_true.columns.values[1] = "Y_TRUE"
y_pred = pd.DataFrame(best_model.predict(x_test, y_test.tolist()),
columns=["Y_PRED"])
confusion_data = pd.merge(left=y_true,
right=y_pred,
left_index=True,
right_index=True)
# summary statistics and metrics
fr_col_nam_map = {0: "feature_nm", 1: "feature_importance"}
feature_ranking = pd.DataFrame(
[X.columns, best_model.feature_importances_]).T
feature_ranking = feature_ranking.rename(columns=fr_col_nam_map)
feature_ranking = feature_ranking.sort_values("feature_nm",
ascending=False)
metrics = {
"precision_score":
precision_score(confusion_data['Y_TRUE'], confusion_data['Y_PRED']),
"roc_auc_score":
roc_auc_score(confusion_data['Y_TRUE'], confusion_data['Y_PRED']),
"classification_report":
classification_report(confusion_data['Y_TRUE'],
confusion_data['Y_PRED']),
"confusion_matrix":
confusion_matrix(confusion_data['Y_TRUE'], confusion_data['Y_PRED']),
"accuracy_score":
accuracy_score(confusion_data['Y_TRUE'], confusion_data['Y_PRED']),
"precision_recall_curve":
precision_recall_curve(confusion_data['Y_TRUE'],
confusion_data['Y_PRED']),
"recall_score":
recall_score(confusion_data['Y_TRUE'], confusion_data['Y_PRED']),
"roc_curve":
roc_curve(confusion_data['Y_TRUE'], confusion_data['Y_PRED'])
}
output = {
"model_name": model_name # string with model name
,
"model_class": best_model # grid_search.best_estimator_
,
"optimized_parameters": optimized_parameters # best_model.get_params()
,
"feature_ranking": feature_ranking # best_model.feature_importances_
,
"metrics": metrics,
"confusion_data": confusion_data
}
return output
