TensorFlow2.0训练模型时，指标不收敛一直上升到1

Question

我尝试着使用tf2.0来搭建一个DeepFM模型来预测用户是否喜欢某部影片，

optimizer选择Adam，loss选择BinaryCrossentropy，评价指标是AUC；

因为涉及到了影片ID，所以我用了shared_embedding，并且必须关闭eager模式；

选用binary_crossentropy作为损失函数时模型在训练时AUC很快就到1了，但选用categorical_crossentropy时loss没太大变化，并且AUC一直保持在0.5，准确率也一直在0.5附近震荡。

下面是选用binary_crossentropy时的输出日志：

下面是我的代码：

one_order_feature_layer = tf.keras.layers.DenseFeatures(one_order_feature_columns)
one_order_feature_layer_outputs = one_order_feature_layer(feature_layer_inputs)

two_order_feature_layer = tf.keras.layers.DenseFeatures(two_order_feature_columns)
two_order_feature_layer_outputs = two_order_feature_layer(feature_layer_inputs)

# lr部分
lr_layer = tf.keras.layers.Dense(len(one_order_feature_columns), kernel_initializer=initializer)(
    one_order_feature_layer_outputs)

# fm部分
reshape = tf.reshape(two_order_feature_layer_outputs,
                     [-1, len(two_order_feature_columns), two_order_feature_columns[0].dimension])
sum_square = tf.square(tf.reduce_sum(reshape, axis=1))
square_sum = tf.reduce_sum(tf.square(reshape), axis=1)
fm_layers = tf.multiply(0.5, tf.subtract(sum_square, square_sum))

# DNN部分
dnn_hidden_layer_1 = tf.keras.layers.Dense(64, activation='selu', kernel_initializer=initializer,
                                           kernel_regularizer=regularizer)(two_order_feature_layer_outputs)
dnn_hidden_layer_2 = tf.keras.layers.Dense(64, activation='selu', kernel_initializer=initializer,
                                           kernel_regularizer=regularizer)(dnn_hidden_layer_1)
dnn_hidden_layer_3 = tf.keras.layers.Dense(64, activation='selu', kernel_initializer=initializer,
                                           kernel_regularizer=regularizer)(dnn_hidden_layer_2)
dnn_dropout = tf.keras.layers.Dropout(0.5, seed=29)(dnn_hidden_layer_3)

# 连接并输出
concatenate_layer = tf.keras.layers.concatenate([lr_layer, fm_layers, dnn_dropout])
out_layer = tf.keras.layers.Dense(1, activation='sigmoid')(concatenate_layer)
model = tf.keras.Model(inputs=[v for v in feature_layer_inputs.values()], outputs=out_layer)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
              loss=tf.keras.losses.BinaryCrossentropy(),
              metrics=['AUC'])
# tf.keras.utils.plot_model(model, 'test.png', show_shapes=True)
train_ds = make_dataset(train_df, buffer_size=None, shuffle=True)
test_ds = make_dataset(test_df)
with tf.compat.v1.Session() as sess:
    sess.run([tf.compat.v1.global_variables_initializer(), tf.compat.v1.tables_initializer()])
    model.fit(train_ds, epochs=5)
    loss, auc = model.evaluate(test_ds)
    print("AUC", auc)

Answer 1

不同的损失函数之间LOSS没有可比性。关键看你的auc（正确率）
因为你是2分类，所以0.5等于没有效果。在数据平衡的前提下，如果达到0.8 0.9就是有效果了。