特征列 通常用於對結構化數據實施特征工程時候使用，圖像或者文本數據一般不會用到特征列。
一，特征列用法概述
使用特征列可以將類別特征轉換為one-hot編碼特征，將連續特征構建分桶特征，以及對多個特征生成交叉特征等等。
要創建特征列，請調用 tf.feature_column 模塊的函數。該模塊中常用的九個函數如下圖所示，所有九個函數都會返回一個 Categorical-Column 或一個 Dense-Column 對象，但卻不會返回 bucketized_column，后者繼承自這兩個類。
注意：所有的Catogorical Column類型最終都要通過indicator_column轉換成Dense Column類型才能傳入模型！
- numeric_column 數值列，最常用。
- bucketized_column 分桶列，由數值列生成，可以由一個數值列出多個特征，one-hot編碼。
- categorical_column_with_identity 分類標識列，one-hot編碼，相當於分桶列每個桶為1個整數的情況。
- categorical_column_with_vocabulary_list 分類詞匯列，one-hot編碼，由list指定詞典。
- categorical_column_with_vocabulary_file 分類詞匯列，由文件file指定詞典。
- categorical_column_with_hash_bucket 哈希列，整數或詞典較大時采用。
- indicator_column 指標列，由Categorical Column生成，one-hot編碼
- embedding_column 嵌入列，由Categorical Column生成，嵌入矢量分布參數需要學習。嵌入矢量維數建議取類別數量的 4 次方根。
- crossed_column 交叉列，可以由除categorical_column_with_hash_bucket的任意分類列構成。
二，特征列使用范例
以下是一個使用特征列解決Titanic生存問題的完整范例。
import datetime import numpy as np import pandas as pd from matplotlib import pyplot as plt import tensorflow as tf from tensorflow.keras import layers,models # 打印日志 def printlog(info): nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S') print("\n"+"=========="*8 + "%s"%nowtime) print(info+'...\n\n') # ================================================================================ # 一，構建數據管道 # ================================================================================ printlog("step1: prepare dataset...") dftrain_raw = pd.read_csv("./data/titanic/train.csv") dftest_raw = pd.read_csv("./data/titanic/test.csv") dfraw = pd.concat([dftrain_raw,dftest_raw]) def prepare_dfdata(dfraw): dfdata = dfraw.copy() dfdata.columns = [x.lower() for x in dfdata.columns] dfdata = dfdata.rename(columns={'survived':'label'}) dfdata = dfdata.drop(['passengerid','name'],axis = 1) for col,dtype in dict(dfdata.dtypes).items(): # 判斷是否包含缺失值 if dfdata[col].hasnans: # 添加標識是否缺失列 dfdata[col + '_nan'] = pd.isna(dfdata[col]).astype('int32') # 填充 if dtype not in [np.object,np.str,np.unicode]: dfdata[col].fillna(dfdata[col].mean(),inplace = True) else: dfdata[col].fillna('',inplace = True) return(dfdata) dfdata = prepare_dfdata(dfraw) dftrain = dfdata.iloc[0:len(dftrain_raw),:] dftest = dfdata.iloc[len(dftrain_raw):,:] # 從 dataframe 導入數據 def df_to_dataset(df, shuffle=True, batch_size=32): dfdata = df.copy() if 'label' not in dfdata.columns: ds = tf.data.Dataset.from_tensor_slices(dfdata.to_dict(orient = 'list')) else: labels = dfdata.pop('label') ds = tf.data.Dataset.from_tensor_slices((dfdata.to_dict(orient = 'list'), labels)) if shuffle: ds = ds.shuffle(buffer_size=len(dfdata)) ds = ds.batch(batch_size) return ds ds_train = df_to_dataset(dftrain) ds_test = df_to_dataset(dftest) # ================================================================================ # 二，定義特征列 # ================================================================================ printlog("step2: make feature columns...") feature_columns = [] # 數值列 for col in ['age','fare','parch','sibsp'] + [ c for c in dfdata.columns if c.endswith('_nan')]: feature_columns.append(tf.feature_column.numeric_column(col)) # 分桶列 age = tf.feature_column.numeric_column('age') age_buckets = tf.feature_column.bucketized_column(age, boundaries=[18, 25, 30, 35, 40, 45, 50, 55, 60, 65]) feature_columns.append(age_buckets) # 類別列 # 注意：所有的Catogorical Column類型最終都要通過indicator_column轉換成Dense Column類型才能傳入模型！！ sex = tf.feature_column.indicator_column( tf.feature_column.categorical_column_with_vocabulary_list( key='sex',vocabulary_list=["male", "female"])) feature_columns.append(sex) pclass = tf.feature_column.indicator_column( tf.feature_column.categorical_column_with_vocabulary_list( key='pclass',vocabulary_list=[1,2,3])) feature_columns.append(pclass) ticket = tf.feature_column.indicator_column( tf.feature_column.categorical_column_with_hash_bucket('ticket',3)) feature_columns.append(ticket) embarked = tf.feature_column.indicator_column( tf.feature_column.categorical_column_with_vocabulary_list( key='embarked',vocabulary_list=['S','C','B'])) feature_columns.append(embarked) # 嵌入列 cabin = tf.feature_column.embedding_column( tf.feature_column.categorical_column_with_hash_bucket('cabin',32),2) feature_columns.append(cabin) # 交叉列 pclass_cate = tf.feature_column.categorical_column_with_vocabulary_list( key='pclass',vocabulary_list=[1,2,3]) crossed_feature = tf.feature_column.indicator_column( tf.feature_column.crossed_column([age_buckets, pclass_cate],hash_bucket_size=15)) feature_columns.append(crossed_feature) # ================================================================================ # 三，定義模型 # ================================================================================ printlog("step3: define model...") tf.keras.backend.clear_session() model = tf.keras.Sequential([ layers.DenseFeatures(feature_columns), #將特征列放入到tf.keras.layers.DenseFeatures中!!! layers.Dense(64, activation='relu'), layers.Dense(64, activation='relu'), layers.Dense(1, activation='sigmoid') ]) # ================================================================================ # 四，訓練模型 # ================================================================================ printlog("step4: train model...") model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy']) history = model.fit(ds_train, validation_data=ds_test, epochs=10) # ================================================================================ # 五，評估模型 # ================================================================================ printlog("step5: eval model...") model.summary() %matplotlib inline %config InlineBackend.figure_format = 'svg' import matplotlib.pyplot as plt def plot_metric(history, metric): train_metrics = history.history[metric] val_metrics = history.history['val_'+metric] epochs = range(1, len(train_metrics) + 1) plt.plot(epochs, train_metrics, 'bo--') plt.plot(epochs, val_metrics, 'ro-') plt.title('Training and validation '+ metric) plt.xlabel("Epochs") plt.ylabel(metric) plt.legend(["train_"+metric, 'val_'+metric]) plt.show() plot_metric(history,"accuracy")
================================================================================2020-04-13 02:29:07 step1: prepare dataset...... ================================================================================2020-04-13 02:29:08 step2: make feature columns...... ================================================================================2020-04-13 02:29:08 step3: define model...... ================================================================================2020-04-13 02:29:08 step4: train model...... Epoch 1/10 23/23 [==============================] - 0s 21ms/step - loss: 0.7117 - accuracy: 0.6615 - val_loss: 0.5706 - val_accuracy: 0.7039 Epoch 2/10 23/23 [==============================] - 0s 3ms/step - loss: 0.5920 - accuracy: 0.7022 - val_loss: 0.6129 - val_accuracy: 0.6648 Epoch 3/10 23/23 [==============================] - 0s 3ms/step - loss: 0.6388 - accuracy: 0.7079 - val_loss: 0.5196 - val_accuracy: 0.7374 Epoch 4/10 23/23 [==============================] - 0s 3ms/step - loss: 0.5950 - accuracy: 0.7219 - val_loss: 0.5028 - val_accuracy: 0.7318 Epoch 5/10 23/23 [==============================] - 0s 3ms/step - loss: 0.5166 - accuracy: 0.7486 - val_loss: 0.4975 - val_accuracy: 0.7318 Epoch 6/10 23/23 [==============================] - 0s 3ms/step - loss: 0.5260 - accuracy: 0.7612 - val_loss: 0.5045 - val_accuracy: 0.7821 Epoch 7/10 23/23 [==============================] - 0s 3ms/step - loss: 0.4957 - accuracy: 0.7697 - val_loss: 0.4756 - val_accuracy: 0.7709 Epoch 8/10 23/23 [==============================] - 0s 3ms/step - loss: 0.4848 - accuracy: 0.7837 - val_loss: 0.4532 - val_accuracy: 0.8045 Epoch 9/10 23/23 [==============================] - 0s 3ms/step - loss: 0.4636 - accuracy: 0.8006 - val_loss: 0.4561 - val_accuracy: 0.7989 Epoch 10/10 23/23 [==============================] - 0s 3ms/step - loss: 0.4784 - accuracy: 0.7907 - val_loss: 0.4722 - val_accuracy: 0.7821 ================================================================================2020-04-13 02:29:11 step5: eval model...... Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= dense_features (DenseFeature multiple 64 _________________________________________________________________ dense (Dense) multiple 3008 _________________________________________________________________ dense_1 (Dense) multiple 4160 _________________________________________________________________ dense_2 (Dense) multiple 65 ================================================================= Total params: 7,297 Trainable params: 7,297 Non-trainable params: 0 _________________________________________________________________
參考：
開源電子書地址：https://lyhue1991.github.io/eat_tensorflow2_in_30_days/
GitHub 項目地址：https://github.com/lyhue1991/eat_tensorflow2_in_30_days