美购精排模型

mlp/train.py

@@ -70,28 +70,30 @@ def getDataSet(df,shuffleSize = 10000,batchSize=128):
     return dataSet
 
 def getTrainColumns(train_columns,data_vocab):
-    columns = []
+    deep_columns = []
+    wide_columns = []
     dataColumns = []
     inputs = {}
     # 离散特征
     for feature in train_columns:
         if data_vocab.get(feature):
-            if feature.count("__")>0 or feature.count("Bucket") > 0:
+            if feature.count("__")>0:
                 cat_col = tf.feature_column.categorical_column_with_vocabulary_list(key=feature,vocabulary_list=data_vocab[feature])
                 col = tf.feature_column.embedding_column(cat_col, 5)
-                columns.append(col)
+                deep_columns.append(col)
                 dataColumns.append(feature)
                 inputs[feature] = tf.keras.layers.Input(name=feature, shape=(), dtype='string')
-            elif feature in one_hot_columns:
+            elif feature in one_hot_columns or feature.count("Bucket") > 0:
                 cat_col = tf.feature_column.categorical_column_with_vocabulary_list(key=feature, vocabulary_list=data_vocab[feature])
-                col = tf.feature_column.indicator_column(cat_col)
-                columns.append(col)
+#                 col = tf.feature_column.indicator_column(cat_col)
+                col = tf.feature_column.embedding_column(cat_col, 3)
+                wide_columns.append(col)
                 dataColumns.append(feature)
                 inputs[feature] = tf.keras.layers.Input(name=feature, shape=(), dtype='string')
             else:
                 cat_col = tf.feature_column.categorical_column_with_vocabulary_list(key=feature,vocabulary_list=data_vocab[feature])
                 col = tf.feature_column.embedding_column(cat_col, 10)
-                columns.append(col)
+                deep_columns.append(col)
                 dataColumns.append(feature)
                 inputs[feature] = tf.keras.layers.Input(name=feature, shape=(), dtype='string')
 #         if feature.startswith("userRatedHistory") or feature.count("__") > 0 or feature in embedding_columns:
@@ -111,21 +113,35 @@ def getTrainColumns(train_columns,data_vocab):
 
         elif feature in ITEM_NUMBER_COLUMNS:
             col = tf.feature_column.numeric_column(feature)
-            columns.append(col)
+            wide_columns.append(col)
             dataColumns.append(feature)
             inputs[feature] = tf.keras.layers.Input(name=feature, shape=(), dtype='float32')
 
 
-    return columns,dataColumns,inputs
+    return deep_columns,wide_columns,dataColumns,inputs
 
 
-def train(columns,inputs,train_dataset):
-    deep = tf.keras.layers.DenseFeatures(columns)(inputs)
-    deep = tf.keras.layers.Dense(64, activation='relu')(deep)
-    deep = tf.keras.layers.Dropout(0.2)(deep)
+def train(deep_columns,wide_columns,inputs,train_dataset):
+    wide = tf.keras.layers.DenseFeatures(deep_columns + wide_columns)(inputs)
+    deep = tf.keras.layers.DenseFeatures(deep_columns)(inputs)
+
     deep = tf.keras.layers.Dense(64, activation='relu')(deep)
     deep = tf.keras.layers.Dropout(0.2)(deep)
-    output_layer = tf.keras.layers.Dense(1, activation='sigmoid')(deep)
+    concat_layer = tf.keras.layers.concatenate([wide, deep], axis=1)
+
+    #     deep = tf.keras.layers.Dense(64, activation='relu')(deep)
+    #     deep = tf.keras.layers.Dropout(0.5)(deep)
+    output_layer = tf.keras.layers.Dense(1, activation='sigmoid')(concat_layer)
+    #     output_layer = FM(1)(deep)
+
+
+    # deep = tf.keras.layers.DenseFeatures(columns)(inputs)
+    # deep = tf.keras.layers.Dense(64, activation='relu')(deep)
+    # deep = tf.keras.layers.Dropout(0.2)(deep)
+    # deep = tf.keras.layers.Dense(64, activation='relu')(deep)
+    # deep = tf.keras.layers.Dropout(0.2)(deep)
+    # output_layer = tf.keras.layers.Dense(1, activation='sigmoid')(deep)
+
     model = tf.keras.Model(inputs, output_layer)
 
     # compile the model, set loss function, optimizer and evaluation metrics
@@ -135,7 +151,7 @@ def train(columns,inputs,train_dataset):
         metrics=['accuracy', tf.keras.metrics.AUC(curve='ROC'), tf.keras.metrics.AUC(curve='PR')])
 
     # train the model
-    model.fit(train_dataset, epochs=5)
+    model.fit(train_dataset, epochs=10)
     print("train save...")
 
     model.save(model_file, include_optimizer=False, save_format='tf')
@@ -183,7 +199,7 @@ if __name__ == '__main__':
 
     # 获取训练列
     columns = df_train.columns.tolist()
-    trainColumns, datasColumns,inputs = getTrainColumns(columns, data_vocab)
+    deep_columns,wide_columns, datasColumns,inputs = getTrainColumns(columns, data_vocab)
 
     df_train = df_train[datasColumns + ["label"]]
     df_test = df_test[datasColumns + ["label"]]
@@ -203,7 +219,7 @@ if __name__ == '__main__':
 
     print("train start...")
     timestmp3 = int(round(time.time()))
-    model = train(trainColumns,inputs,train_data)
+    model = train(deep_columns,wide_columns,inputs,train_data)
     timestmp4 = int(round(time.time()))
     print("train end...耗时h:{}".format((timestmp4 - timestmp3)/60/60))