Wide&Deep——tf 代码

1、分类
@estimator_export('estimator.DNNLinearCombinedClassifier')
class DNNLinearCombinedClassifier(estimator.Estimator):Example:

```python
numeric_feature = numeric_column(...)
categorical_column_a = categorical_column_with_hash_bucket(...)
categorical_column_b = categorical_column_with_hash_bucket(...)

categorical_feature_a_x_categorical_feature_b = crossed_column(...)
categorical_feature_a_emb = embedding_column(
    categorical_column=categorical_feature_a, ...)
categorical_feature_b_emb = embedding_column(
    categorical_id_column=categorical_feature_b, ...)

estimator = DNNLinearCombinedClassifier(
    # wide settings
    linear_feature_columns=[categorical_feature_a_x_categorical_feature_b],
    linear_optimizer=tf.train.FtrlOptimizer(...),
    # deep settings
    dnn_feature_columns=[
        categorical_feature_a_emb, categorical_feature_b_emb,
        numeric_feature],
    dnn_hidden_units=[1000, 500, 100],
    dnn_optimizer=tf.train.ProximalAdagradOptimizer(...),
    # warm-start settings
    warm_start_from="/path/to/checkpoint/dir")

# To apply L1 and L2 regularization, you can set dnn_optimizer to:
tf.train.ProximalAdagradOptimizer(
    learning_rate=0.1,
    l1_regularization_strength=0.001,
    l2_regularization_strength=0.001)
# To apply learning rate decay, you can set dnn_optimizer to a callable:
lambda: tf.AdamOptimizer(
    learning_rate=tf.exponential_decay(
        learning_rate=0.1,
        global_step=tf.get_global_step(),
        decay_steps=10000,
        decay_rate=0.96)
# It is the same for linear_optimizer.

# Input builders
def input_fn_train: # returns x, y
  pass
estimator.train(input_fn=input_fn_train, steps=100)

def input_fn_eval: # returns x, y
  pass
metrics = estimator.evaluate(input_fn=input_fn_eval, steps=10)
def input_fn_predict: # returns x, None
  pass
predictions = estimator.predict(input_fn=input_fn_predict)

```

在‘train’和‘evaluate’时，需注意：
`dnn_feature_columns` + `linear_feature_columns`中的‘columns’需要满足以下条件：
column如果是‘_CategoricalColumn’，feature的key=column.name，value是一个‘SparseTensor’;
column如果是‘_WeightedCategoricalColumn’，两个feature，第一个的key是id column name，第二个的key是weight column name，两个特征的value必须是一个SparseTensor;
column如果是‘_DenseColumn’，特征的key=column.name，value是一个Tensor

Loss是用softmax cross entropy计算的。

2、回归
@estimator_export('estimator.DNNLinearCombinedRegressor')
class DNNLinearCombinedRegressor(estimator.Estimator):Example:

```python
numeric_feature = numeric_column(...)
categorical_column_a = categorical_column_with_hash_bucket(...)
categorical_column_b = categorical_column_with_hash_bucket(...)

categorical_feature_a_x_categorical_feature_b = crossed_column(...)
categorical_feature_a_emb = embedding_column(
    categorical_column=categorical_feature_a, ...)
categorical_feature_b_emb = embedding_column(
    categorical_column=categorical_feature_b, ...)

estimator = DNNLinearCombinedRegressor(
    # wide settings
    linear_feature_columns=[categorical_feature_a_x_categorical_feature_b],
    linear_optimizer=tf.train.FtrlOptimizer(...),
    # deep settings
    dnn_feature_columns=[
        categorical_feature_a_emb, categorical_feature_b_emb,
        numeric_feature],
    dnn_hidden_units=[1000, 500, 100],
    dnn_optimizer=tf.train.ProximalAdagradOptimizer(...),
    # warm-start settings
    warm_start_from="/path/to/checkpoint/dir")

# To apply L1 and L2 regularization, you can set dnn_optimizer to:
tf.train.ProximalAdagradOptimizer(
    learning_rate=0.1,
    l1_regularization_strength=0.001,
    l2_regularization_strength=0.001)
# To apply learning rate decay, you can set dnn_optimizer to a callable:
lambda: tf.AdamOptimizer(
    learning_rate=tf.exponential_decay(
        learning_rate=0.1,
        global_step=tf.get_global_step(),
        decay_steps=10000,
        decay_rate=0.96)
# It is the same for linear_optimizer.

# Input builders
def input_fn_train: # returns x, y
  pass
estimator.train(input_fn=input_fn_train, steps=100)

def input_fn_eval: # returns x, y
  pass
metrics = estimator.evaluate(input_fn=input_fn_eval, steps=10)
def input_fn_predict: # returns x, None
  pass
predictions = estimator.predict(input_fn=input_fn_predict)
```

在‘train’和‘evaluate’时，需注意：
`dnn_feature_columns` + `linear_feature_columns`中的‘columns’需要满足以下条件：
column如果是‘_CategoricalColumn’，feature的key=column.name，value是一个‘SparseTensor’;
column如果是‘_WeightedCategoricalColumn’，两个feature，第一个的key是id column name，第二个的key是weight column name，两个特征的value必须是一个SparseTensor;
column如果是‘_DenseColumn’，特征的key=column.name，value是一个Tensor

Loss是用mean squared error计算的。

 

 

3、特征列

线性部分：任何类型的特征列

深度部分：密集列特征，其他的列需要用indicator_column、embedding_column封装

 

4、区别

 

5、主函数

def _dnn_linear_combined_model_fn(features,
                                  labels,
                                  mode,
                                  head,
                                  linear_feature_columns=None,
                                  linear_optimizer='Ftrl',
                                  dnn_feature_columns=None,
                                  dnn_optimizer='Adagrad',
                                  dnn_hidden_units=None,
                                  dnn_activation_fn=nn.relu,
                                  dnn_dropout=None,
                                  input_layer_partitioner=None,
                                  config=None,
                                  batch_norm=False,
                                  linear_sparse_combiner='sum'):