代码
print("[INFO] build mlp module")
#queries [B, D]
queries = input_layer[0]
#keys/facts shape=[B, sequence_length, embedding_size]
keys = input_layer[1]
# get dimension of query
queries_hidden_units = queries.get_shape().as_list()[-1]
# tile queries for sequence length， [B, queries_D] -> [B, queries_D * sequence_length]
print("tf.shape(keys):", keys.get_shape())
print("tf.shape(queries):", queries.get_shape())
queries = tf.tile(queries, [1, tf.shape(keys)[1]])
# [B, sequence_length, embedding_size]
# query and fact are of the same dimension
queries = tf.reshape(queries,
[-1, tf.shape(keys)[1], queries_hidden_units])
if len(input_layer) == 3:
timegap = input_layer[2] # [B, sequence_length, d]
din_all = tf.concat([queries, keys, queries - keys, queries * keys, timegap],
axis=-1)
else:
din_all = tf.concat([queries, keys, queries - keys, queries * keys],
axis=-1)
print("din attention mlp input shape: ", din_all.get_shape())
# units = 80, 输出的维度大小，改变inputs的最后一维
# 即在 [B, sequence_length, concat_embedding] 的最后一维度 变成 80
# mlp:[80, 40, 1]
d_layer_1_all = tf.layers.dense(din_all,
80,
activation=tf.nn.sigmoid,
name="{0}_f1_att".format(self.name),
reuse=tf.AUTO_REUSE)
d_layer_2_all = tf.layers.dense(
d_layer_1_all,
40,
activation=tf.nn.sigmoid,
name="{0}_f2_att".format(self.name),
reuse=tf.AUTO_REUSE,
)
d_layer_3_all = tf.layers.dense(d_layer_2_all,
1,
activation=None,
name="{0}_f3_att".format(self.name),
reuse=tf.AUTO_REUSE)
d_layer_3_all = tf.reshape(d_layer_3_all, [-1, 1, tf.shape(keys)[1]])
outputs = d_layer_3_all
# Mask
key_masks = tf.sequence_mask(
#lenghths = [B,B,B,B,B,B,B...B]
# maxlen = T, 控制最大长度
tf.ones(
shape=tf.shape(keys)[0],
dtype=tf.int32) * tf.shape(keys)[1], tf.shape(keys)[1]) # [B, T]
# [B, 1, T] all true
key_masks = tf.expand_dims(key_masks, 1) # [B, 1, T]
paddings = tf.ones_like(outputs) * (-(2**32) + 1)
outputs = tf.where(key_masks, outputs, paddings) # [B, 1, T]
# Scale
outputs = outputs / (keys.get_shape().as_list()[-1]**0.5)
# Activation
outputs = tf.nn.softmax(outputs) # [B, 1, T]
# Weighted sum
# [B, 1, T] * [B, T, D]
output = tf.matmul(outputs, keys) # [B, 1, D]
# tf.squeeze 与 tf.expand_dims互为逆操作
output = tf.squeeze(output, [1])
return output
输入部分
- Keys: 表示待查询的key，其结构为[B, Sequence_length, Embedding_size]
- B: 表示Batch Size的大小
- Sequence_length: 表示序列的长度
- Embedding_size:表示序列嵌入的大小
- Query: 表示查询的Qurey，在DIN中表示Target Item的Embedding。其结构为[B, Sequence_length, D]
- B: 表示Batch Size的大小。
- Sequence_length:表示序列的长度。
- D: 表示Query序列的Embedding长度。
这里先不考虑Time Gap的操作。
操作部分
- tf.tile(queries, [1, tf.shape(keys)[1]):
这里执行了一个tile操作, 我们来看一下tile操作的功能是什么。
这里假设Batch Size = 3 sequence_length = 5 embedding_size = 4 Queries_D = 6
import tensorflow as tf
a = [[[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0]], [[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0]], [[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0],
[1.0, 2.0, 3.0, 4.0]]]
b = [[1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
[1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
[1.0, 2.0, 3.0, 4.0, 5.0, 6.0]]
keys = tf.convert_to_tensor(a)
queries = tf.convert_to_tensor(b)
queries_hidden_units = queries.get_shape().as_list()[-1] #获取embedding size
print(tf.shape(keys)[1])
print(queries)
queries = tf.tile(queries, [1,tf.shape(keys)[1]]) # 在第一维上不扩增，在第二维度上扩增embedding size的大小
结果输出如下:
tf.Tensor(5, shape=(), dtype=int32)
tf.Tensor(
[[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]], shape=(3, 6), dtype=float32)
tf.Tensor(
[[1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6.
1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6.
1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6. 1. 2. 3. 4. 5. 6.
1. 2. 3. 4. 5. 6.]], shape=(3, 30), dtype=float32)
- 将Queries 转换成[B, Sequence_length, queries_D]的大小
queries = tf.reshape(queries,
[-1, tf.shape(keys)[1], queries_hidden_units])
print(queries)
tf.Tensor(
[[[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]]
[[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]]
[[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]
[1. 2. 3. 4. 5. 6.]]], shape=(3, 5, 6), dtype=float32)