einsum教程网站Einstein Summation in Numpy | Olexa Bilaniuk's IFT6266H16 Course Blog
编写训练模型
import tensorflow as tf
class Model(tf.keras.Model):
def __init__(self, num_heads, model_dim):
super().__init__()
self.AttentionLayer = tf.keras.layers.MultiHeadAttention(num_heads=num_heads, key_dim=model_dim)
self.OutputLayer = tf.keras.layers.Dense(units=1)
def call(self, x):
x = self.AttentionLayer(query=x, value=x)
x = self.OutputLayer(x)
return x
model = Model(num_heads=2, model_dim=4)
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
optimizer="Adam",
metrics=['accuracy'])
input_train = tf.constant([[[1, 2, 3], [4, 5, 6]],
[[1, 1, 1], [2, 2, 6]]], dtype=tf.float32)
output_label = tf.constant([[[1], [0]],
[[0], [0]]], dtype=tf.float32)
tf_callback = tf.keras.callbacks.TensorBoard(log_dir="./logs")
model.fit(x=input_train,
y=output_label,
epochs=10,
callbacks=[tf_callback])
tf.saved_model.save(model, 'MultiHeadAttention')
训练样本input_train的shape为(2, 2, 3)→(batch_size, sentence_length, Embedding_dim),经过模型后的输出shape为(2,2,1),标签值的shape为(2, 2, 1),损失函数选择的是二分类交叉熵。所以该模型可以应用于二分类性质的命名实体识别。简单来说,该模型可以判断一个句子中哪些词属于我们感兴趣的词类,哪些不属于。
打印模型矩阵参数
import tensorflow as tf
save_path = 'MultiHeadAttention/variables/variables' #
reader = tf.train.load_checkpoint(save_path) # 得到CheckpointReader
""" 打印Checkpoint中存储的所有参数名和参数shape """
for variable_name, variable_shape in reader.get_variable_to_shape_map().items():
print(f'{variable_name} : {variable_shape}')
打印结果,
_CHECKPOINTABLE_OBJECT_GRAPH : []
optimizer/beta_1/.ATTRIBUTES/VARIABLE_VALUE : []
keras_api/metrics/0/count/.ATTRIBUTES/VARIABLE_VALUE : []
variables/5/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
optimizer/beta_2/.ATTRIBUTES/VARIABLE_VALUE : []
optimizer/learning_rate/.ATTRIBUTES/VARIABLE_VALUE : []
keras_api/metrics/0/total/.ATTRIBUTES/VARIABLE_VALUE : []
keras_api/metrics/1/count/.ATTRIBUTES/VARIABLE_VALUE : []
keras_api/metrics/1/total/.ATTRIBUTES/VARIABLE_VALUE : []
optimizer/decay/.ATTRIBUTES/VARIABLE_VALUE : []
variables/2/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
optimizer/iter/.ATTRIBUTES/VARIABLE_VALUE : []
variables/0/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/0/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/0/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/1/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/1/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/5/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/1/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/6/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [2, 4, 3]
variables/2/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/2/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/3/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/3/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/3/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/4/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/4/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/4/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/5/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/9/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [1]
variables/6/.ATTRIBUTES/VARIABLE_VALUE : [2, 4, 3]
variables/6/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [2, 4, 3]
variables/7/.ATTRIBUTES/VARIABLE_VALUE : [3]
variables/8/.ATTRIBUTES/VARIABLE_VALUE : [3, 1]
variables/7/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [3]
variables/7/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [3]
variables/8/.OPTIMIZER_SLOT/optimizer/m/.ATTRIBUTES/VARIABLE_VALUE : [3, 1]
variables/8/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [3, 1]
variables/9/.ATTRIBUTES/VARIABLE_VALUE : [1]
variables/9/.OPTIMIZER_SLOT/optimizer/v/.ATTRIBUTES/VARIABLE_VALUE : [1]
其中我们只需要注意如下参数,
variables/0/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/1/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/5/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/2/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/3/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]
variables/4/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]
variables/6/.ATTRIBUTES/VARIABLE_VALUE : [2, 4, 3]
variables/7/.ATTRIBUTES/VARIABLE_VALUE : [3]
variables/8/.ATTRIBUTES/VARIABLE_VALUE : [3, 1]variables/9/.ATTRIBUTES/VARIABLE_VALUE : [1]
这些参数代表着模型中所有的kernel和bias,variable后面的序号代表着输入在模型中依次经历的运算。利用模型内部的计算逻辑以及shape分布,我们可以找出对应的kernel和bias。
值得注意的是,在tf.keras.layers.MultiHeadAttention源码中,query映射层是先进行计算的,后面依次为key映射层,value映射层。所以,
variables/0/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4] 对应query映射层的kernel
variables/1/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]对应query映射层的bias
variables/2/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]对应key映射层的kernel
variables/3/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]对应key映射层的bias
variables/4/.ATTRIBUTES/VARIABLE_VALUE : [3, 2, 4]对应value映射层的kernel
variables/5/.ATTRIBUTES/VARIABLE_VALUE : [2, 4]对应value映射层的bias
variables/6/.ATTRIBUTES/VARIABLE_VALUE : [2, 4, 3]对应MutiHeadAttention中输出映射层的kernel
variables/7/.ATTRIBUTES/VARIABLE_VALUE : [3]对应MutiHeadAttention中输出映射层的bias
variables/8/.ATTRIBUTES/VARIABLE_VALUE : [3, 1]对应Dense层的kernel
variables/9/.ATTRIBUTES/VARIABLE_VALUE : [1]对应Dense层的bias
MultiHeadAttention
多头映射层
''' abc , cde -> abde
a : size of batch
b : length of sequence
c : dimension of Embedding
d : number of heads
e : dimension of output
x : input, shape=(batch_size, sequence_length, dim)
y : kenel, shape=(dim, heads_num, output_dim) '''
def project_dense(x, y):
output = np.empty((x.shape[0], x.shape[1], y.shape[1], y.shape[2]))
for a in range(x.shape[0]):
for b in range(x.shape[1]):
for d in range(y.shape[1]):
for e in range(y.shape[2]):
sum_dot = 0
for c in range(x.shape[2]):
sum_dot += x[a][b][c] * y[c][d][e]
output[a][b][d][e] = sum_dot
return output
代码验证,
''' shape = (2, 2, 3) '''
x = tf.constant([[[1, 2, 3], [4, 5, 3]],
[[7, 8, 9], [10, 11, 12]]])
''' shape = (3, 2, 3) '''
y = tf.constant([[[1, 2, 3], [4, 5, 3]],
[[7, 8, 9], [10, 11, 12]],
[[7, 8, 9], [10, 11, 12]]])
print(tf.einsum('abc,cde->abde', x, y).numpy())
print(project_dense(x, y))
[[[[ 36 42 48]
[ 54 60 63]]
[[ 60 72 84]
[ 96 108 108]]]
[[[126 150 174]
[198 222 225]]
[[171 204 237]
[270 303 306]]]]
[[[[ 36. 42. 48.]
[ 54. 60. 63.]]
[[ 60. 72. 84.]
[ 96. 108. 108.]]]
[[[126. 150. 174.]
[198. 222. 225.]]
[[171. 204. 237.]
[270. 303. 306.]]]]
多头注意力层
计算注意力分数矩阵
''' aecd, abcd -> acbe
a : size of batch
e : number of key
b : number of query
c : number of heads
d : dimension of query/key
x is key and y is query '''
def compute_AttentionScores(x, y):
output = np.empty((x.shape[0], x.shape[2], y.shape[1], x.shape[1]))
for a in range(x.shape[0]):
for c in range(x.shape[2]):
for b in range(y.shape[1]):
for e in range(x.shape[1]):
sum_dot = 0
for d in range(x.shape[3]):
sum_dot += x[a][e][c][d] * y[a][b][c][d]
output[a][c][b][e] = sum_dot
return output
根据注意力分数对value进行加权叠加
''' acbe,aecd->abcd
a : size of batch
c : number of heads
b : number of query
e : number of key/value
d : dimension of value
x is attention_scores and y is value'''
def Value_WeightedStack(x, y):
output = np.empty((x.shape[0], x.shape[2], y.shape[2], y.shape[3]))
for a in range(x.shape[0]):
for b in range(x.shape[2]):
for c in range(y.shape[2]):
for d in range(y.shape[3]):
sum_dot = 0
for e in range(x.shape[3]):
sum_dot += x[a][c][b][e] * y[a][e][c][d]
output[a][b][c][d] = sum_dot
return output
输出映射层
''' abcd, cde -> abe
a : size of batch
b : number of query
c : number of head
d : dimension of value
e : dimension of output
x is WeightedStack_Value and y is kernel'''
def project_final(x, y):
output = np.empty((x.shape[0], x.shape[1], y.shape[2]))
for a in range(x.shape[0]):
for b in range(x.shape[1]):
for e in range(y.shape[2]):
sum_dot = 0
for c in range(x.shape[2]):
for d in range(x.shape[3]):
sum_dot += x[a][b][c][d] * y[c][d][e]
output[a][b][e] = sum_dot
return output
Dense
为了使得方便验证这次试验,在MultiHeadAttentionn后面添加一个Dense层。
'''abc, cd -> abd
x is input and y is kernel '''
def output_dense(x, y):
output = np.empty((x.shape[0], x.shape[1], y.shape[1]))
for a in range(x.shape[0]):
for b in range(x.shape[1]):
for d in range(y.shape[1]):
sum_dot = 0
for c in range(x.shape[2]):
sum_dot += x[a][b][c] * y[c][d]
output[a][b][d] = sum_dot
return output
构建模型的前向传播
在前面我们找出了模型内部所有的kenel和bias,接下来我们将打印出这些参数,并将这些参数添加到我们自己编写的模型中去,实现前向传播。
print(reader.get_tensor('variables/0/.ATTRIBUTES/VARIABLE_VALUE')) //k1
print(reader.get_tensor('variables/1/.ATTRIBUTES/VARIABLE_VALUE')) //b1
print(reader.get_tensor('variables/5/.ATTRIBUTES/VARIABLE_VALUE')) //b2
print(reader.get_tensor('variables/2/.ATTRIBUTES/VARIABLE_VALUE')) //k2
print(reader.get_tensor('variables/3/.ATTRIBUTES/VARIABLE_VALUE')) //b3
print(reader.get_tensor('variables/4/.ATTRIBUTES/VARIABLE_VALUE')) //k3
print(reader.get_tensor('variables/6/.ATTRIBUTES/VARIABLE_VALUE')) //MultiHead_output_kerner
print(reader.get_tensor('variables/7/.ATTRIBUTES/VARIABLE_VALUE')) //MultiHead_output_bias
print(reader.get_tensor('variables/8/.ATTRIBUTES/VARIABLE_VALUE')) //output_dense_kernel
print(reader.get_tensor('variables/9/.ATTRIBUTES/VARIABLE_VALUE')) //output_dense_bias
class my_model:
def __init__(self, input):
self.input = input
def __call__(self):
x = tf.cast(self.input,dtype=tf.double)
value = tf.add(project_dense(x, k3), b2)
key = tf.add(project_dense(x, k2), b3)
query = tf.add(project_dense(x, k1), b1)
attention_scores = tf.nn.softmax(compute_AttentionScores(key, query), axis=-1)
Stacked_value = Value_WeightedStack(attention_scores, value)
MUltiHead_output = tf.add(project_final(Stacked_value, MultiHead_output_kerner), MultiHead_output_bias)
output = tf.add(output_dense(MUltiHead_output, output_dense_kernel), output_dense_bias)
return output
验证
最后我们将分别打印出两个模型前向传播的输出(一个是自己实现的,另一个是TensorFlow实现的),并进行结果比对,看是否相差无几。
test_in = tf.constant([[[0.1, 0.1, 0.1], [0.1, 0.1, 0.1]]], dtype=tf.float32)
test_in0 = tf.constant([[[1, 1, 1], [1, 1, 1]]], dtype=tf.float32)
test_in1 = tf.constant([[[10, 10, 10], [10, 10, 10]]], dtype=tf.float32)
test_in2 = tf.constant([[[100, 100, 100], [100, 100, 100]]], dtype=tf.float32)
test_in3 = tf.constant([[[1000, 1000, 1000], [1000, 1000, 1000]]], dtype=tf.float32)
model = tf.saved_model.load('MultiHeadAttention')
print(model(test_in))
print(model(test_in0))
print(model(test_in1))
print(model(test_in2))
print(model(test_in3))
tf.Tensor(
[[[-0.02398136]
[-0.02398136]]], shape=(1, 2, 1), dtype=float32)
tf.Tensor(
[[[-0.75980777]
[-0.75980777]]], shape=(1, 2, 1), dtype=float32)
tf.Tensor(
[[[-8.1180725]
[-8.1180725]]], shape=(1, 2, 1), dtype=float32)
tf.Tensor(
[[[-81.700714]
[-81.700714]]], shape=(1, 2, 1), dtype=float32)
tf.Tensor(
[[[-817.52716]
[-817.52716]]], shape=(1, 2, 1), dtype=float32)
test_in = tf.constant([[[0.1, 0.1, 0.1], [0.1, 0.1, 0.1]]], dtype=tf.float32)
test_in0 = tf.constant([[[1, 1, 1], [1, 1, 1]]], dtype=tf.float32)
test_in1 = tf.constant([[[10, 10, 10], [10, 10, 10]]], dtype=tf.float32)
test_in2 = tf.constant([[[100, 100, 100], [100, 100, 100]]], dtype=tf.float32)
test_in3 = tf.constant([[[1000, 1000, 1000], [1000, 1000, 1000]]], dtype=tf.float32)
print(my_model(test_in)())
print(my_model(test_in0)())
print(my_model(test_in1)())
print(my_model(test_in2)())
print(my_model(test_in3)())
tf.Tensor(
[[[-0.02398137]
[-0.02398137]]], shape=(1, 2, 1), dtype=float64)
tf.Tensor(
[[[-0.75980776]
[-0.75980776]]], shape=(1, 2, 1), dtype=float64)
tf.Tensor(
[[[-8.11807168]
[-8.11807168]]], shape=(1, 2, 1), dtype=float64)
tf.Tensor(
[[[-81.70071091]
[-81.70071091]]], shape=(1, 2, 1), dtype=float64)
tf.Tensor(
[[[-817.5271032]
[-817.5271032]]], shape=(1, 2, 1), dtype=float64)
两个输出结果相差无几,验证成功。
此外值得注意的是,在整个MultiHeadAttention计算机制当中并没有出现激活函数。