踩坑內容包含以下 1. feature_column的輸入輸出型別，用一個數據集給出demo 2. feature_column接estimator 3. feature_column接Keras ## feature_column 輸入輸出型別 ### 輸入輸出型別 feature_column輸入可以是原始特徵的列名,或者是feature_column。初上手感覺feature_column設計的有點奇怪，不過熟悉了邏輯後用起來還是很方便的。幾個需要習慣一下的點： 1. 深度模型的輸入必須是Dense型別，所有輸出是categorical型別需要經過indicator或者embedding的轉換才可以 2. indicator, embedding, bucketized的輸入不能是原始特徵，前兩者只能是categorical型別的feature_column, 後者只能是numeric_column |feature_column| 輸入| 輸出|輸出是否為dense| |----|----|----|---| |categorical_column_with_identity|數值型離散|categorical|N| |categorical_column_with_vocabulary_list|字元型/數值型離散|categorical|N| |categorical_column_with_hash_bucket|類別太多的離散值|categorical|N| |crossed_column|categorical/離散值 |categorical|N| |indicator_column|categorical|one/multi-hot|Y| |embedding_column |categorical|dense vector|Y| |numeric_column|數值型連續值|numeric|Y| |bucketzied_column|numeric_column|one-hot|Y| 以下給出各種特徵工程的demo，原始特徵如下 ![image.png-252.2kB][1] ### 輸入-連續值 ![image.png-170.8kB][2] ### 輸入-離散值 ![image.png-286.2kB][3] ### 輸入-categorical ![image.png-290kB][4] ## feature_column接estimator 如果是使用預定義的estimator, feature_column可以直接作為輸入，不需要任何額外操作，只需要注意深度模型只支援Dense型別的feature_column即可。 如果是自定義estimator，則需要多一步用feature_column先建立input_layer ```python input_layer = tf.feature_column.input_layer(features, feature_columns) ``` ## feature_column接keras 為什麼要這麼搭配呢，好像是沒啥必要，只不過進一步證明tf的官方文件確實坑而已。。。 ```python def model_fn(): #define Keras input input = {} for f in FEATURE_NAME: input[f] = Input(shape=(1,), name = f, dtype = DTYPE[f]) #generate feature_columns feature_columns = build_features() #Define transformation from feature_columns to Dense Tensor feature_layer = tf.keras.layers.DenseFeatures( feature_columns ) #Transform input dense_feature = feature_layer(input) output = Dense(1, activation='sigmoid')(dense_feature) #feed input placeholder as list model = Model(inputs = [i for i in input.values()], outputs = output) return model ``` [1]: http://static.zybuluo.com/hongchenzimo/iyjk0m2i1i631jv9gwj8jzgg/image.png [2]: http://static.zybuluo.com/hongchenzimo/y3ihfhb465aci4n0qs9kvqag/image.png [3]: http://static.zybuluo.com/hongchenzimo/qyfvqwib5turc587l3hml196/image.png [4]: http://static.zybuluo.com/hongchenzimo/6kjyawpwbsdrqjcak96hw3fk/i