在Pytorch中屏蔽Tensorflow中的零填充嵌入（和返回零梯度）

我正在尝试使用Tensorflow中的BPR loss从Spotlight重新创建PoolNet但是我可以使用（它是估算器model_fn） .

def _pooling_model_fn(features, labels, mode, params):
 with tf.name_scope('inputs'):
    if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
        users_prev_items_inputs_train = features['item_seqs']
    elif mode == tf.estimator.ModeKeys.PREDICT:
        users_prev_items_inputs_train = tf.reshape(features['item_seqs'], [1, -1])

 with tf.device('/cpu:0'):
    prod_embeddings = tf.keras.layers.Embedding(params["num_items"], params["item_emb_size"], mask_zero=True)
    item_biases = tf.keras.layers.Embedding(params["num_items"], 1, mask_zero=True, embeddings_initializer=tf.keras.initializers.Zeros())
    prod_embed = prod_embeddings(users_prev_items_inputs_train)
    targets = tf.transpose(prod_embed, [0, 2, 1])

 sequence_embeddings = tf.expand_dims(targets, axis=3)

 sequence_embeddings = tf.pad(sequence_embeddings, paddings=tf.constant([[0, 0], [0, 0], [1, 0], [0, 0]]))

 sequence_embedding_sum = tf.cumsum(sequence_embeddings, 2)

 non_padding_entries = tf.cumsum(tf.cast(tf.not_equal(sequence_embeddings, tf.constant(0.0)), tf.float32), 2)  # .expand_as(sequence_embedding_sum)

 user_representations = tf.squeeze((sequence_embedding_sum / (non_padding_entries + 1)), [3])

 user_representations_so_far = user_representations[:, :, :-1]
 user_representations_new = user_representations[:, :, -1]

 if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
    global_step = tf.contrib.framework.get_or_create_global_step()

    with tf.name_scope('loss'):
        negative_samples = features['neg_samp']

        with tf.device('/cpu:0'):
            prod_embed_pos = prod_embeddings(users_prev_items_inputs_train)
            target_embedding_positive = tf.squeeze(tf.transpose(prod_embed_pos, [0, 2, 1]))

            prod_bias_pos = item_biases(users_prev_items_inputs_train) 
            target_bias_positive = tf.squeeze(prod_bias_pos)

        dot_positive = tf.reduce_sum(user_representations_so_far * target_embedding_positive, 1) + target_bias_positive

        with tf.device('/cpu:0'):
            prod_embed_neg = prod_embeddings(negative_samples)
            target_embedding_negative = tf.squeeze(tf.transpose(prod_embed_neg, [0, 2, 1]))

            prod_bias_neg = item_biases(negative_samples)
            target_bias_negative = tf.squeeze(prod_bias_neg)

        dot_negative = tf.reduce_sum(user_representations_so_far * target_embedding_negative, 1) + target_bias_negative

        mask = tf.not_equal(users_prev_items_inputs_train, 0)

        loss = bpr_loss(dot_positive, dot_negative, mask)

 if mode == tf.estimator.ModeKeys.TRAIN:
    with tf.name_scope('optimizer'):
        optimizer = tf.train.AdamOptimizer(learning_rate=params["lr"])
    train_op = optimizer.minimize(loss, global_step=global_step)
    return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)

 if mode == tf.estimator.ModeKeys.PREDICT:
    item_ids = np.arange(params['num_items']).reshape(-1, 1)
    item_ids_tensor = tf.convert_to_tensor(item_ids, dtype=tf.int64)

    with tf.device('/cpu:0'):
        prod_embed_pos = prod_embeddings(item_ids_tensor)  # tf.nn.embedding_lookup(prod_embeddings, item_ids_tensor)
        target_embedding_positive = tf.squeeze(tf.transpose(prod_embed_pos, [0, 2, 1]))

        prod_bias_pos = item_biases(item_ids_tensor)  # tf.nn.embedding_lookup(item_biases, item_ids_tensor)
        target_bias_positive = tf.squeeze(prod_bias_pos)

    dot_positive = tf.reduce_sum(user_representations_new * target_embedding_positive, 1) + target_bias_positive

    predictions = {
        'products': tf.reshape(dot_positive, [1, -1])
    }
    export_outputs = {
        'prediction': tf.estimator.export.PredictOutput(predictions)
    }
    return tf.estimator.EstimatorSpec(mode, predictions=predictions, export_outputs=export_outputs)

和损失功能

def bpr_loss(positive_predictions, negative_predictions, mask):
 loss1 = 1.0 - tf.nn.sigmoid(positive_predictions - negative_predictions)

 if mask is not None:
    mask = tf.cast(mask, loss1.dtype)
    final_loss = loss1 * mask
    return tf.reduce_sum(final_loss) / tf.reduce_sum(mask)

 return tf.reduce_mean(loss1)

使用上述模型，我无法对与Spotlight完全相同的数据集（和相同的随机种子）进行相同的预测 . 我最终认为问题在于零填充 . 生成数据的方式如下：

[[0,0,0,5,6,98],
 [0,62,15,4,8,47],
 [0,0,5,9,6,3,41],
 [78,21,2,56,1,3]]

它们具有前导零填充，因此每个输入样本具有相同的长度 .

基于我的代码，我相信我做了一切来掩盖丢失，嵌入层（使用Keras的mask_zero参数）以及我正在计算的嵌入（使用cumsum）的平均值中的这些零 . 尽管如此，在训练之后，零索引嵌入也在不断变化（意味着不考虑排除，而是导致影响其余梯度并为我的结果添加噪声） .

Pytorch在Embedding layer的实现中似乎有一个很好的功能，你可以设置 padding_idx 与pad的id，这将用零初始化 . 此外，它使该索引的梯度始终为零 . 基本上，我试图用Tensorflow做同样的事情 .

任何帮助，将不胜感激 .