ESMM模型预估cvr全为0

meilejw commented 4 years ago

hi 请教个问题 ESMM模型预估的ctr看起来正常但是预估cvr全为0 会是什么原因呢？

jiwenbiao1 commented 4 years ago

请问哪里有数据集，直接可以用的

fenxouxiaoquan commented 4 years ago

hi 请教个问题 ESMM模型预估的ctr看起来正常但是预估cvr全为0 会是什么原因呢？

我也遇到这个问题了，目前已解决，方式是：build_mode()函数只返回net，把计算logits的步骤放在my_model()函数里搞，

def my_model(features, labels, mode, params):
    with tf.variable_scope('ctr_model'):
        last_ctr_layer = build_mode(features, mode, params)
    with tf.variable_scope('cvr_model'):
        last_cvr_layer = build_mode(features, mode, params)

    head = head_lib._binary_logistic_or_multi_class_head(
            n_classes=2, weight_column=None, label_vocabulary=None,
            loss_reduction=losses.Reduction.SUM)

    ctr_logits = tf.layers.dense(last_ctr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())
    cvr_logits = tf.layers.dense(last_cvr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())

    ctr_predictions = tf.sigmoid(ctr_logits, name="CTR")
    cvr_predictions = tf.sigmoid(cvr_logits, name="CVR")
    .........

ywc1026 commented 3 years ago

hi 请教个问题 ESMM模型预估的ctr看起来正常但是预估cvr全为0 会是什么原因呢？

我也遇到这个问题了，目前已解决，方式是：build_mode()函数只返回net，把计算logits的步骤放在my_model()函数里搞，

def my_model(features, labels, mode, params):
    with tf.variable_scope('ctr_model'):
        last_ctr_layer = build_mode(features, mode, params)
    with tf.variable_scope('cvr_model'):
        last_cvr_layer = build_mode(features, mode, params)

    head = head_lib._binary_logistic_or_multi_class_head(
            n_classes=2, weight_column=None, label_vocabulary=None,
            loss_reduction=losses.Reduction.SUM)

    ctr_logits = tf.layers.dense(last_ctr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())
    cvr_logits = tf.layers.dense(last_cvr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())

    ctr_predictions = tf.sigmoid(ctr_logits, name="CTR")
    cvr_predictions = tf.sigmoid(cvr_logits, name="CVR")
    .........

我试了一下，CVR 还是 0

ywc1026 commented 3 years ago

可能输入里没有 tfrecords 文件。把 train_input_fn 的函数第一句改成 dataset = tf.data.TFRecordDataset(filenames)，然后 filename 里包含 tfrecords ，比如”../../cvr_train.tfrecords“ ，cvr 结果就不为 0

shuDaoNan9 commented 3 years ago

可能输入里没有 tfrecords 文件。把 train_input_fn 的函数第一句改成 dataset = tf.data.TFRecordDataset(filenames)，然后 filename 里包含 tfrecords ，比如”../../cvr_train.tfrecords“ ，cvr 结果就不为 0

你有数据集吗？

ywc1026 commented 3 years ago

可能输入里没有 tfrecords 文件。把 train_input_fn 的函数第一句改成 dataset = tf.data.TFRecordDataset(filenames)，然后 filename 里包含 tfrecords ，比如”../../cvr_train.tfrecords“ ，cvr 结果就不为 0

你有数据集吗？

我用的自己的数据集

copypastejun commented 2 years ago

hi 请教个问题 ESMM模型预估的ctr看起来正常但是预估cvr全为0 会是什么原因呢？

我也遇到这个问题了，目前已解决，方式是：build_mode()函数只返回net，把计算logits的步骤放在my_model()函数里搞，
def my_model(features, labels, mode, params):
    with tf.variable_scope('ctr_model'):
        last_ctr_layer = build_mode(features, mode, params)
    with tf.variable_scope('cvr_model'):
        last_cvr_layer = build_mode(features, mode, params)

    head = head_lib._binary_logistic_or_multi_class_head(
            n_classes=2, weight_column=None, label_vocabulary=None,
            loss_reduction=losses.Reduction.SUM)

    ctr_logits = tf.layers.dense(last_ctr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())
    cvr_logits = tf.layers.dense(last_cvr_layer, units=head.logits_dimension,
                                 kernel_initializer=tf.glorot_uniform_initializer())

    ctr_predictions = tf.sigmoid(ctr_logits, name="CTR")
    cvr_predictions = tf.sigmoid(cvr_logits, name="CVR")
    .........
ctr_logits = tf.layers.dense(last_ctr_layer, units=head.logits_dimension, kernel_initializer=tf.glorot_uniform_initializer()) cvr_logits = tf.layers.dense(last_cvr_layer, units=head.logits_dimension, kernel_initializer=tf.glorot_uniform_initializer())

其中 head.logits_dimension 在做二分类的情况下返回的就是 1，效果和 logits = tf.layers.dense(net, 1, activation=None)是一样的啊

@property def logits_dimension(self): return 1

yangxudong / deeplearning

ESMM模型预估cvr全为0 #16