Wide and Deep Learning（2）tf代码实现
https://zhuanlan.zhihu.com/p/43328492
1、数据
人口普查数据：
查看data：
特征进行处理：
连续的特征，包含这些：
CONTINUOUS_COLUMNS = ["age", "education_num", "capital_gain", "capital_loss",
"hours_per_week"]
离散的特征，包含这些：
CATEGORICAL_COLUMNS = ["workclass", "education", "marital_status", "occupation",
"relationship", "race", "gender", "native_country"]
linear模型中
1.离散：
（1）按照自定义的字典将类别特征映射到数值，适合特征种类较少时候使用
gender = tf.contrib.layers.sparse_column_with_keys(column_name=“gender”, keys=[“female”,“male”])
（2）自动将类别特征映射到数值，适合特征种类较多时候使用
education = tf.contrib.layers.sparse_column_with_hash_bucket(“education”, hash_bucket_size = 1000)
2.连续：
（3） 连续数值特征
age = tf.contrib.layers.real_valued_column(“age”)
（4）把连续特征按照区间映射为类别特征
age_buckets = tf.contrib.layers.bucketized_column(age, boundaries= [18,25, 30, 35, 40, 45, 50, 55, 60, 65])
3.交叉特征：
（5）#特征相乘生成的交叉特征
tf.contrib.layers.crossed_column([education, occupation], hash_bucket_size=int(1e4))
神经网络中
1.连续的特征可以直接传入神经网络
2.离散的特征需要做处理之后再传入：有两种方式进行处理，可以使用one-hot 编码处理神经网络输入的离散特征，也可以使用 embedding 处理；建议该特征的数值种类较少时候使用one-hot 编码，在该特征的数值种类较多的情况下使用 embedding 编码，这里之所以要这么处理，我理解的原因是独热编码如果数值种类很多，会导致编码后的特征维度太高，从而学习到的信息过于分散，因此，在种类很多的时候，需要先将超高维的种类离散特征进行压缩embedding,再送入神经网络；
这是把离散特征embedding 操作的核心，它把原始的类别数值映射到这个权值矩阵，其实相当于神经网络的权值，后续如果是trainable的话，我们就会把这个当做网络的权值矩阵进行训练，但是在用的时候，就把这个当成一个embedding表，按id去取每个特征的embedding 后的数值。（这其实就类似于词向量了，把每个单词映射到一个词向量。）
embedding编码：（这其实就类似于词向量了，把每个单词映射到一个词向量。）
tf.contrib.layers.embedding_column(workclass, dimension=8)
离散值的存储：
什么是 sparseTensor？ 什么时候使用这种数据类型？
它是用（位置，值，形状） 三个元素简略的表示一个矩阵的方法，例如：
SparseTensor(indices=[[0, 0], [1, 2]], values=[1, 2], dense_shape=[3, 4])
表示：
也就是说它表明，在（0，0）的位置有1， （1，2）的位置存在2；
它只针对类别特征，适合于表示矩阵中大量元素为0的情况，会大大减少矩阵占用的存储量；在案例中，我发现它经常用来作为一个类别特征的中间变量矩阵，用来减小内存占用：
在input_fn（）中：
categorical_cols = {k: tf.SparseTensor(indices=[[i,0] for i in range( df[k].size)], values = df[k].values, dense_shape=[df[k].size,1]) for k in CATEGORICAL_COLUMNS}
对离散特征的存储。
二、model：
根据选择的model_type不同，使用不同的方法。
if FLAGS.model_type =="wide":
m = tf.contrib.learn.LinearClassifier(model_dir=model_dir,feature_columns=wide_columns)
elif FLAGS.model_type == "deep":
m = tf.contrib.learn.DNNClassifier(model_dir=model_dir, feature_columns=deep_columns, hidden_units=[100,50])
else:
m = tf.contrib.learn.DNNLinearCombinedClassifier(model_dir=model_dir, linear_feature_columns=wide_columns, dnn_feature_columns = deep_columns, dnn_hidden_units=[100,50])
三、整体代码：
1.定义不同的FLAGS选择和离散项，连续项，Label
#tf定义了tf.app.flags，用于支持接受命令行传递参数，相当于接受argv。
import tempfile
import tensorflow as tf
from six.moves import urllib
import pandas as pd
flags = tf.app.flags #tf定义了tf.app.flags，用于支持接受命令行传递参数，相当于接受argv。
FLAGS = flags.FLAGS
#参数的几种选择：
flags.DEFINE_string("model_dir","","Base directory for output models.")
flags.DEFINE_string("model_type","wide_n_deep","valid model types:{'wide','deep', 'wide_n_deep'")
flags.DEFINE_integer("train_steps",200,"Number of training steps.")
flags.DEFINE_string("train_data","", "Path to the training data.")
flags.DEFINE_string("test_data", "", "path to the test data")
COLUMNS = ["age", "workclass", "fnlwgt", "education", "education_num",
"marital_status", "occupation", "relationship", "race", "gender",
"capital_gain", "capital_loss", "hours_per_week", "native_country",
"income_bracket"]
LABEL_COLUMN = "label"
CATEGORICAL_COLUMNS = ["workclass", "education", "marital_status", "occupation",
"relationship", "race", "gender", "native_country"]
CONTINUOUS_COLUMNS = ["age", "education_num", "capital_gain", "capital_loss",
"hours_per_week"]
2.下载数据：maybe_download()
def maybe_download():
if FLAGS.train_data:
train_data_file = FLAGS.train_data
else:
train_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve("http://mlr.cs.umass.edu/ml/machine-learning-databases/adult/adult.data", train_file.name)
train_file_name = train_file.name
train_file.close()
print("Training data is downloaded to %s" % train_file_name)
if FLAGS.test_data:
test_file_name = FLAGS.test_data
else:
test_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve("http://mlr.cs.umass.edu/ml/machine-learning-databases/adult/adult.test",
test_file.name) # pylint: disable=line-too-long
test_file_name = test_file.name
test_file.close()
print("Test data is downloaded to %s" % test_file_name)
return train_file_name, test_file_name
3.定义训练模型：build_estimator（）
但这里线性模型和神经网络的特征采集：
wide 模型的特征都是离散特征、 离散特征之间的交互作用特征；
deep 模型的特征则是离散特征embedding 加上连续特征；
wide 端模型和 deep 端模型只需要分别专注于擅长的方面，wide 端模型通过离散特征的交叉组合进行 memorization，deep 端模型通过特征的 embedding 进行 generalization，这样单个模型的大小和复杂度也能得到控制，而整体模型的性能仍能得到提高。
# build the estimator
def build_estimator(model_dir):
# 离散分类别的
gender = tf.contrib.layers.sparse_column_with_keys(column_name="gender", keys=["female","male"])#按照自定义的字典将类别特征映射到数值，适合特征种类较少时候使用
education = tf.contrib.layers.sparse_column_with_hash_bucket("education", hash_bucket_size = 1000)#自动将类别特征映射到数值，适合特征种类较多时候使用
relationship = tf.contrib.layers.sparse_column_with_hash_bucket("relationship", hash_bucket_size = 100)
workclass = tf.contrib.layers.sparse_column_with_hash_bucket("workclass", hash_bucket_size=100)
occupation = tf.contrib.layers.sparse_column_with_hash_bucket("occupation", hash_bucket_size=1000)
native_country = tf.contrib.layers.sparse_column_with_hash_bucket( "native_country", hash_bucket_size=1000)
# Continuous base columns.
age = tf.contrib.layers.real_valued_column("age")#连续数值特征
education_num = tf.contrib.layers.real_valued_column("education_num")
capital_gain = tf.contrib.layers.real_valued_column("capital_gain")
capital_loss = tf.contrib.layers.real_valued_column("capital_loss")
hours_per_week = tf.contrib.layers.real_valued_column("hours_per_week")
#类别转换
age_buckets = tf.contrib.layers.bucketized_column(age, boundaries= [18,25, 30, 35, 40, 45, 50, 55, 60, 65]) #把连续特征按照区间映射为类别特征
wide_columns = [gender, native_country,education, occupation, workclass, relationship, age_buckets,
tf.contrib.layers.crossed_column([education, occupation], hash_bucket_size=int(1e4)), #特征相乘生成的交叉特征
tf.contrib.layers.crossed_column([age_buckets, education, occupation], hash_bucket_size=int(1e6)),
tf.contrib.layers.crossed_column([native_country, occupation],hash_bucket_size=int(1e4))]
#embedding_column用来表示类别型的变量
#这是把离散特征embedding 操作的核心，它把原始的类别数值映射到这个权值矩阵，其实相当于神经网络的权值，后续如果是trainable的话，我们就会把这个当做网络的权值矩阵进行训练，但是在用的时候，就把这个当成一个embedding表，按id去取每个特征的embedding 后的数值。
# （这其实就类似于词向量了，把每个单词映射到一个词向量。）
deep_columns = [tf.contrib.layers.embedding_column(workclass, dimension=8),
tf.contrib.layers.embedding_column(education, dimension=8),
tf.contrib.layers.embedding_column(gender, dimension=8),
tf.contrib.layers.embedding_column(relationship, dimension=8),
tf.contrib.layers.embedding_column(native_country,dimension=8),
tf.contrib.layers.embedding_column(occupation, dimension=8),
age,education_num,capital_gain,capital_loss,hours_per_week,]
if FLAGS.model_type =="wide":
m = tf.contrib.learn.LinearClassifier(model_dir=model_dir,feature_columns=wide_columns)
elif FLAGS.model_type == "deep":
m = tf.contrib.learn.DNNClassifier(model_dir=model_dir, feature_columns=deep_columns, hidden_units=[100,50])
else:
m = tf.contrib.learn.DNNLinearCombinedClassifier(model_dir=model_dir, linear_feature_columns=wide_columns, dnn_feature_columns = deep_columns, dnn_hidden_units=[100,50])
return m
两个模型使用不同的优化器是什么？ 怎么在loss层面配合？
以 tf.estimator.DNNLinearCombinedClassifier() 为例，进入该函数发现：linear_optimizer=‘Ftrl’, 线性模型使用’Ftrl’优化器，dnn_optimizer=‘Adagrad’, 神经网络使用’Adagrad’优化器；
如果分类数量是2，loss使用
_binary_logistic_head_with_sigmoid_cross_entropy_loss
如果分类数量大于2，loss使用
_multi_class_head_with_softmax_cross_entropy_loss
最后怎么得到两个模型结合的预测结果的？
直接把两个模型的结果相加：
logits = dnn_logits + linear_logits
很出乎意料的方式啊，居然是直接相加的，然后是上面一点提到的操作，用这个logits 和label 得到 Loss，再将 Loss 分别反传回两个独立的优化器分别优化两个模型的权重。
4.划分数据中的Label 和 feature
#划分数据中的Label 和 feature
def input_fn(df):
continuous_cols = {k: tf.constant(df[k].values) for k in CONTINUOUS_COLUMNS}
categorical_cols = {k: tf.SparseTensor(indices=[[i,0] for i in range( df[k].size)], values = df[k].values, dense_shape=[df[k].size,1]) for k in CATEGORICAL_COLUMNS}#原文例子为dense_shape
feature_cols = dict(continuous_cols)
feature_cols.update(categorical_cols)
label = tf.constant(df[LABEL_COLUMN].values)
return feature_cols, label
5.下载文件，读取文件，去掉值为NAN的项，定义Label，划分Label，创建临时目录model_dir
，训练模型，估计测试集，输出最后的指标。
def train_and_eval():
train_file_name, test_file_name = maybe_download()
df_train = pd.read_csv(
tf.gfile.Open(train_file_name),
names=COLUMNS,
skipinitialspace=True,
engine="python"
)
df_test = pd.read_csv(
tf.gfile.Open(test_file_name),
names=COLUMNS,
skipinitialspace=True,
skiprows=1,
engine="python"
)
# drop Not a number elements
df_train = df_train.dropna(how='any',axis=0)
df_test = df_test.dropna(how='any', axis=0)
#convert >50 to 1
df_train[LABEL_COLUMN] = (
df_train["income_bracket"].apply(lambda x: ">50" in x).astype(int)
)
df_test[LABEL_COLUMN] = (
df_test["income_bracket"].apply(lambda x: ">50K" in x)).astype(int)
model_dir = tempfile.mkdtemp() if not FLAGS.model_dir else FLAGS.model_dir
print("model dir = %s" % model_dir)
m = build_estimator(model_dir)
print (FLAGS.train_steps)
m.fit(input_fn=lambda: input_fn(df_train),
steps=FLAGS.train_steps)
results = m.evaluate(input_fn=lambda: input_fn(df_test), steps=1)
for key in sorted(results):
print("%s: %s"%(key, results[key]))
def main(_):
train_and_eval()
if __name__ == "__main__":
tf.app.run()
最后代码为：
import tempfile
import tensorflow as tf
from six.moves import urllib
import pandas as pd
flags = tf.app.flags #tf定义了tf.app.flags，用于支持接受命令行传递参数，相当于接受argv。
FLAGS = flags.FLAGS
#参数的几种选择：
flags.DEFINE_string("model_dir","","Base directory for output models.")
flags.DEFINE_string("model_type","wide_n_deep","valid model types:{'wide','deep', 'wide_n_deep'")
flags.DEFINE_integer("train_steps",200,"Number of training steps.")
flags.DEFINE_string("train_data","", "Path to the training data.")
flags.DEFINE_string("test_data", "", "path to the test data")
COLUMNS = ["age", "workclass", "fnlwgt", "education", "education_num",
"marital_status", "occupation", "relationship", "race", "gender",
"capital_gain", "capital_loss", "hours_per_week", "native_country",
"income_bracket"]
LABEL_COLUMN = "label"
CATEGORICAL_COLUMNS = ["workclass", "education", "marital_status", "occupation",
"relationship", "race", "gender", "native_country"]
CONTINUOUS_COLUMNS = ["age", "education_num", "capital_gain", "capital_loss",
"hours_per_week"]
# download test and train data
def maybe_download():
if FLAGS.train_data:
train_data_file = FLAGS.train_data
else:
train_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve("http://mlr.cs.umass.edu/ml/machine-learning-databases/adult/adult.data", train_file.name)
train_file_name = train_file.name
train_file.close()
print("Training data is downloaded to %s" % train_file_name)
if FLAGS.test_data:
test_file_name = FLAGS.test_data
else:
test_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve("http://mlr.cs.umass.edu/ml/machine-learning-databases/adult/adult.test",
test_file.name) # pylint: disable=line-too-long
test_file_name = test_file.name
test_file.close()
print("Test data is downloaded to %s" % test_file_name)
return train_file_name, test_file_name
# build the estimator
def build_estimator(model_dir):
# 离散分类别的
gender = tf.contrib.layers.sparse_column_with_keys(column_name="gender", keys=["female","male"])#按照自定义的字典将类别特征映射到数值，适合特征种类较少时候使用
education = tf.contrib.layers.sparse_column_with_hash_bucket("education", hash_bucket_size = 1000)#自动将类别特征映射到数值，适合特征种类较多时候使用
relationship = tf.contrib.layers.sparse_column_with_hash_bucket("relationship", hash_bucket_size = 100)
workclass = tf.contrib.layers.sparse_column_with_hash_bucket("workclass", hash_bucket_size=100)
occupation = tf.contrib.layers.sparse_column_with_hash_bucket("occupation", hash_bucket_size=1000)
native_country = tf.contrib.layers.sparse_column_with_hash_bucket( "native_country", hash_bucket_size=1000)
# Continuous base columns.
age = tf.contrib.layers.real_valued_column("age")#连续数值特征
education_num = tf.contrib.layers.real_valued_column("education_num")
capital_gain = tf.contrib.layers.real_valued_column("capital_gain")
capital_loss = tf.contrib.layers.real_valued_column("capital_loss")
hours_per_week = tf.contrib.layers.real_valued_column("hours_per_week")
#类别转换
age_buckets = tf.contrib.layers.bucketized_column(age, boundaries= [18,25, 30, 35, 40, 45, 50, 55, 60, 65]) #把连续特征按照区间映射为类别特征
wide_columns = [gender, native_country,education, occupation, workclass, relationship, age_buckets,
tf.contrib.layers.crossed_column([education, occupation], hash_bucket_size=int(1e4)), #特征相乘生成的交叉特征
tf.contrib.layers.crossed_column([age_buckets, education, occupation], hash_bucket_size=int(1e6)),
tf.contrib.layers.crossed_column([native_country, occupation],hash_bucket_size=int(1e4))]
#embedding_column用来表示类别型的变量
#这是把离散特征embedding 操作的核心，它把原始的类别数值映射到这个权值矩阵，其实相当于神经网络的权值，后续如果是trainable的话，我们就会把这个当做网络的权值矩阵进行训练，但是在用的时候，就把这个当成一个embedding表，按id去取每个特征的embedding 后的数值。
# （这其实就类似于词向量了，把每个单词映射到一个词向量。）
deep_columns = [tf.contrib.layers.embedding_column(workclass, dimension=8),
tf.contrib.layers.embedding_column(education, dimension=8),
tf.contrib.layers.embedding_column(gender, dimension=8),
tf.contrib.layers.embedding_column(relationship, dimension=8),
tf.contrib.layers.embedding_column(native_country,dimension=8),
tf.contrib.layers.embedding_column(occupation, dimension=8),
age,education_num,capital_gain,capital_loss,hours_per_week,]
if FLAGS.model_type =="wide":
m = tf.contrib.learn.LinearClassifier(model_dir=model_dir,feature_columns=wide_columns)
elif FLAGS.model_type == "deep":
m = tf.contrib.learn.DNNClassifier(model_dir=model_dir, feature_columns=deep_columns, hidden_units=[100,50])
else:
m = tf.contrib.learn.DNNLinearCombinedClassifier(model_dir=model_dir, linear_feature_columns=wide_columns, dnn_feature_columns = deep_columns, dnn_hidden_units=[100,50])
return m
#划分数据中的Label 和 feature
def input_fn(df):
continuous_cols = {k: tf.constant(df[k].values) for k in CONTINUOUS_COLUMNS}
categorical_cols = {k: tf.SparseTensor(indices=[[i,0] for i in range( df[k].size)], values = df[k].values, dense_shape=[df[k].size,1]) for k in CATEGORICAL_COLUMNS}#原文例子为dense_shape
feature_cols = dict(continuous_cols)
feature_cols.update(categorical_cols)
label = tf.constant(df[LABEL_COLUMN].values)
return feature_cols, label
#训练的主函数
def train_and_eval():
train_file_name, test_file_name = maybe_download()
df_train = pd.read_csv(
tf.gfile.Open(train_file_name),
names=COLUMNS,
skipinitialspace=True,
engine="python"
)
df_test = pd.read_csv(
tf.gfile.Open(test_file_name),
names=COLUMNS,
skipinitialspace=True,
skiprows=1,
engine="python"
)
# drop Not a number elements
df_train = df_train.dropna(how='any',axis=0)
df_test = df_test.dropna(how='any', axis=0)
#convert >50 to 1
df_train[LABEL_COLUMN] = (
df_train["income_bracket"].apply(lambda x: ">50" in x).astype(int)
)
df_test[LABEL_COLUMN] = (
df_test["income_bracket"].apply(lambda x: ">50K" in x)).astype(int)
model_dir = tempfile.mkdtemp() if not FLAGS.model_dir else FLAGS.model_dir #创建临时目录，目录需要手动删除。
print("model dir = %s" % model_dir)
m = build_estimator(model_dir)
print (FLAGS.train_steps)
m.fit(input_fn=lambda: input_fn(df_train),
steps=FLAGS.train_steps)
results = m.evaluate(input_fn=lambda: input_fn(df_test), steps=1)
for key in sorted(results):
print("%s: %s"%(key, results[key]))
def main(_):
train_and_eval()
if __name__ == "__main__":
tf.app.run()