def call(self, inputs):
if not hasattr(self, 'kernel'):
embedding_layer = inputs._keras_history[0]
if embedding_layer.name != self.embedding_name:
def recursive_search(layer):
"""递归向上搜索，根据名字找Embedding层
"""
last_layer = layer._inbound_nodes[0].inbound_layers
if isinstance(last_layer, list):
if len(last_layer) == 0:
return None
else:
last_layer = last_layer[0]
if last_layer.name == self.embedding_name:
return last_layer
else:
return recursive_search(last_layer)
embedding_layer = recursive_search(embedding_layer)
if embedding_layer is None:
raise Exception('Embedding layer not found')
self.kernel = K.transpose(embedding_layer.embeddings)
self.units = K.int_shape(self.kernel)[1]
self.bias = self.add_weight(name='bias',
shape=(self.units, ),
initializer='zeros')
outputs = K.dot(inputs, self.kernel)
outputs = K.bias_add(outputs, self.bias)
outputs = self.activation(outputs)
return outputs
def call(self, inputs):
ndim, shape = K.ndim(inputs), K.shape(inputs)
shape = [shape[i] for i in range(ndim)]
inputs = K.reshape(
inputs, shape[:-1] + [shape[-1] // self.groups, self.groups])
outputs = tf.einsum('...ig,ijg->...gj', inputs, self.kernel)
outputs = K.reshape(outputs, shape[:-1] + [self.units])
if self.use_bias:
outputs = K.bias_add(outputs, self.bias)
outputs = self.activation(outputs)
return outputs
def call(self, inputs):
if not hasattr(self, 'kernel'):
embedding_layer = search_layer(inputs, self.embedding_name)
if embedding_layer is None:
raise Exception('Embedding layer not found')
self.kernel = K.transpose(embedding_layer.embeddings)
self.units = K.int_shape(self.kernel)[1]
self.bias = self.add_weight(name='bias',
shape=(self.units, ),
initializer='zeros')
outputs = K.dot(inputs, self.kernel)
outputs = K.bias_add(outputs, self.bias)
outputs = self.activation(outputs)
return outputs
def call(self, x):
x, mask = x
mask = K.squeeze(mask, axis=2)
# linear
key = K.bias_add(K.dot(x, self.weight), self.bias)
# compute attention
outputs = K.squeeze(K.dot(key, self.query), axis=2)
outputs -= 1e32 * (1 - mask)
attn_scores = K.softmax(outputs)
attn_scores *= mask
attn_scores = K.reshape(attn_scores,
shape=(-1, 1, attn_scores.shape[-1]))
outputs = K.squeeze(K.batch_dot(attn_scores, key), axis=1)
return outputs
def call(self, x):
x, mask = x
# 因为 self.weight只有两个维度，所以这里要进行维度处理
mask = K.squeeze(mask, axis=2) # 维度压缩 去掉一个维度 axis=2,但是数据还是不变的
# linear 线性变化
# K.dot()进行 点乘，然后加了self.bias
key = K.bias_add(K.dot(x, self.weight), self.bias)
# compute attention
outputs = K.squeeze(K.dot(key, self.query), axis=2) # 计算注意力
outputs -= 1e32 * (1 - mask)
attn_scores = K.softmax(outputs) # 使用 softmax 计算得分
attn_scores *= mask
attn_scores = K.reshape(attn_scores,
shape=(-1, 1, attn_scores.shape[-1]))
outputs = K.squeeze(K.batch_dot(attn_scores, key), axis=1)
return outputs
def call(self, inputs):
return K.bias_add(inputs, self.bias)
