H2O学习笔记（四）——Deep Learning 例子

这里写图片描述##MNIST Digit Classification

MNIST一个比较出名的数据集，包括60000张训练图片和10000张测试图片，每张图片是一个手写数字，包括28²像素值的手写识别数据,扫描的手写数字如下图所示：

Example in python

import h2o
from h2o.estimators.deeplearning import H2ODeepLearningEstimator
# Start H2O cluster with all available cores (default)
h2o.init()
# Get a brief summary of the data
train.describe()
test.describe()

train = h2o.import_file("./train.csv.gz")
test = h2o.import_file("./test.csv.gz")
x=train.names[0:784]
y="C785"


# Encode the response column as categorical for multinomial classification
train[y] = train[y].asfactor()
test[y] = test[y].asfactor()


model = H2ODeepLearningEstimator(distribution='multinomial',
                               activation='RectifierWithDropout',
                               hidden=[32,32,32],
                               input_dropout_ratio=0.2,
                               sparse=True,
                               l1=1e-5,
                               epochs=10,
                               nfolds=5)

model.train(x=x,y=y,training_frame=train,validation_frame=test)
print model.params
print model.model_performance(train=True)
print model.model_performance(test=True)
print model.auc(train=True)
print model.predict(test_data=test)
# Get a brief summary of the data
# train.describe()
# test.describe()

Cartesian Grid Search

对超参数hidden和l1两个参数做网格搜索

from h2o.estimators import H2ODeepLearningEstimator
from h2o.grid.grid_search import H2OGridSearch
import h2o


h2o.init()
train = h2o.import_file("./train.csv.gz")
test = h2o.import_file("./test.csv.gz")

# Get a brief summary of the data
train.describe()
test.describe()

# Specify the response and predictor columns
x=train.names[0:784]
y="C785"
# Encode the response column as categorical for multinomial classification
train[y] = train[y].asfactor()
test[y] = test[y].asfactor()

hidden_opt = [[32,32],[32,16,8],[100]]
l1_opt = [1e-4,1e-3]
hyper_parameters = {"hidden":hidden_opt, "l1":l1_opt}

model_grid = H2OGridSearch(H2ODeepLearningEstimator,
    hyper_params=hyper_parameters)
model_grid.train(x=x, y=y,
                 distribution="multinomial",
                 epochs=1000,
                 training_frame=train,
                 validation_frame=test,
                 score_interval=2,
                 stopping_rounds=3,
                 stopping_tolerance=0.05,
                 stopping_metric="misclassification")

#print model grid search results
for model in model_grid:
    print model.model_id + " mse: " + str(model.mse())

搜索结果：

Random Grid Search

如果参数的搜索空间太大，可以使用随机网格搜索，只需要指定最大的模型个数和seed,使用search_criteria来设置

hidden_opt =[[17,32],[8,19],[32,16,8],[100],[10,10,10,10]]
l1_opt = [s/1e6 for s in range(1,1001)]
hyper_parameters = {"hidden":hidden_opt, "l1":l1_opt}
search_criteria = {"strategy":"RandomDiscrete",
                   "max_models":10,
                   "max_runtime_secs":100,
                   "seed":123456}

model_grid = H2OGridSearch(H2ODeepLearningEstimator,
                        hyper_params=hyper_parameters,
                        search_criteria = search_criteria)
model_grid.train(x=x, y=y,
                 distribution="multinomial",
                 epochs=1000,
                 training_frame=train,
                 validation_frame=test,
                 score_interval=2,
                 stopping_rounds=3,
                 stopping_tolerance=0.05,
                 stopping_metric="misclassification")

#print model grid search results
for model in model_grid:
    print model.model_id + " mse: " + str(model.mse())

搜索结果：