[深度应用]·Keras极简实现Attention结构
1.生成数据
这里让x[:, attention_column] = y[:, 0]，X数据的第一列等于Y数据第零列（其实就是label），这样第一列数据和label的相关度就会很大，最后通过输出相关度来证明思路正确性。
import keras.backend as K
import numpy as np
def get_activations(model, inputs, print_shape_only=False, layer_name=None):
# Documentation is available online on Github at the address below.
# From: https://github.com/philipperemy/keras-visualize-activations
print('----- activations -----')
activations = []
inp = model.input
if layer_name is None:
outputs = [layer.output for layer in model.layers]
else:
outputs = [layer.output for layer in model.layers if layer.name == layer_name] # all layer outputs
funcs = [K.function([inp] + [K.learning_phase()], [out]) for out in outputs] # evaluation functions
layer_outputs = [func([inputs, 1.])[0] for func in funcs]
for layer_activations in layer_outputs:
activations.append(layer_activations)
if print_shape_only:
print(layer_activations.shape)
else:
print(layer_activations)
return activations
def get_data(n, input_dim, attention_column=1):
"""
Data generation. x is purely random except that it's first value equals the target y.
In practice, the network should learn that the target = x[attention_column].
Therefore, most of its attention should be focused on the value addressed by attention_column.
:param n: the number of samples to retrieve.
:param input_dim: the number of dimensions of each element in the series.
:param attention_column: the column linked to the target. Everything else is purely random.
:return: x: model inputs, y: model targets
"""
x = np.random.standard_normal(size=(n, input_dim))
y = np.random.randint(low=0, high=2, size=(n, 1))
x[:, attention_column] = y[:, 0]
return x, y
2.定义网络
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
from attention_utils import get_activations, get_data
np.random.seed(1337) # for reproducibility
from keras.models import *
from keras.layers import Input, Dense,Multiply,Activation
input_dim = 4
def Att(att_dim,inputs,name):
V = inputs
QK = Dense(att_dim,bias=None)(inputs)
QK = Activation("softmax",name=name)(QK)
MV = Multiply()([V, QK])
return(MV)
def build_model():
inputs = Input(shape=(input_dim,))
atts1 = Att(input_dim,inputs,"attention_vec")
x = Dense(16)(atts1)
atts2 = Att(16,x,"attention_vec1")
output = Dense(1, activation='sigmoid')(atts2)
model = Model(input=inputs, output=output)
return model3.训练与作图
if __name__ == '__main__':
N = 10000
inputs_1, outputs = get_data(N, input_dim)
print(inputs_1[:2],outputs[:2])
m = build_model()
m.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
print(m.summary())
m.fit(inputs_1, outputs, epochs=20, batch_size=128, validation_split=0.2)
testing_inputs_1, testing_outputs = get_data(1, input_dim)
# Attention vector corresponds to the second matrix.
# The first one is the Inputs output.
attention_vector = get_activations(m, testing_inputs_1,
print_shape_only=True,
layer_name='attention_vec')[0].flatten()
print('attention =', attention_vector)
# plot part.
pd.DataFrame(attention_vector, columns=['attention (%)']).plot(kind='bar',
title='Attention Mechanism as '
'a function of input'
' dimensions.')
plt.show()4.结果展示
实验结果表明，第一列相关性最大，符合最初的思想。
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 4) 0
__________________________________________________________________________________________________
dense_1 (Dense) (None, 4) 16 input_1[0][0]
__________________________________________________________________________________________________
attention_vec (Activation) (None, 4) 0 dense_1[0][0]
__________________________________________________________________________________________________
multiply_1 (Multiply) (None, 4) 0 input_1[0][0]
attention_vec[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 16) 80 multiply_1[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) (None, 16) 256 dense_2[0][0]
__________________________________________________________________________________________________
attention_vec1 (Activation) (None, 16) 0 dense_3[0][0]
__________________________________________________________________________________________________
multiply_2 (Multiply) (None, 16) 0 dense_2[0][0]
attention_vec1[0][0]
__________________________________________________________________________________________________
dense_4 (Dense) (None, 1) 17 multiply_2[0][0]
==================================================================================================
Total params: 369
Trainable params: 369
Non-trainable params: 0
__________________________________________________________________________________________________
None
Train on 8000 samples, validate on 2000 samples
Epoch 1/20
2019-05-26 20:02:22.289119: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2019-05-26 20:02:22.290211: I tensorflow/core/common_runtime/process_util.cc:69] Creating new thread pool with default inter op setting: 4. Tune using inter_op_parallelism_threads for best performance.
8000/8000 [==============================] - 2s 188us/step - loss: 0.6918 - acc: 0.5938 - val_loss: 0.6893 - val_acc: 0.7715
Epoch 2/20
8000/8000 [==============================] - 0s 23us/step - loss: 0.6848 - acc: 0.7889 - val_loss: 0.6774 - val_acc: 0.8065
Epoch 3/20
8000/8000 [==============================] - 0s 28us/step - loss: 0.6619 - acc: 0.8091 - val_loss: 0.6417 - val_acc: 0.7780
Epoch 4/20
8000/8000 [==============================] - 0s 29us/step - loss: 0.6132 - acc: 0.8166 - val_loss: 0.5771 - val_acc: 0.8610
Epoch 5/20
8000/8000 [==============================] - 0s 28us/step - loss: 0.5304 - acc: 0.8925 - val_loss: 0.4758 - val_acc: 0.9185
Epoch 6/20
8000/8000 [==============================] - 0s 28us/step - loss: 0.4177 - acc: 0.9433 - val_loss: 0.3554 - val_acc: 0.9680
Epoch 7/20
8000/8000 [==============================] - 0s 24us/step - loss: 0.3028 - acc: 0.9824 - val_loss: 0.2533 - val_acc: 0.9930
Epoch 8/20
8000/8000 [==============================] - 0s 40us/step - loss: 0.2180 - acc: 0.9961 - val_loss: 0.1872 - val_acc: 0.9985
Epoch 9/20
8000/8000 [==============================] - 0s 37us/step - loss: 0.1634 - acc: 0.9986 - val_loss: 0.1442 - val_acc: 0.9985
Epoch 10/20
8000/8000 [==============================] - 0s 33us/step - loss: 0.1269 - acc: 0.9998 - val_loss: 0.1140 - val_acc: 0.9985
Epoch 11/20
8000/8000 [==============================] - 0s 22us/step - loss: 0.1013 - acc: 0.9998 - val_loss: 0.0921 - val_acc: 0.9990
Epoch 12/20
8000/8000 [==============================] - 0s 28us/step - loss: 0.0825 - acc: 0.9999 - val_loss: 0.0758 - val_acc: 0.9995
Epoch 13/20
8000/8000 [==============================] - 0s 22us/step - loss: 0.0682 - acc: 1.0000 - val_loss: 0.0636 - val_acc: 0.9995
Epoch 14/20
8000/8000 [==============================] - 0s 20us/step - loss: 0.0572 - acc: 0.9999 - val_loss: 0.0538 - val_acc: 0.9995
Epoch 15/20
8000/8000 [==============================] - 0s 23us/step - loss: 0.0485 - acc: 1.0000 - val_loss: 0.0460 - val_acc: 0.9995
Epoch 16/20
8000/8000 [==============================] - 0s 22us/step - loss: 0.0416 - acc: 1.0000 - val_loss: 0.0397 - val_acc: 0.9995
Epoch 17/20
8000/8000 [==============================] - 0s 23us/step - loss: 0.0360 - acc: 1.0000 - val_loss: 0.0345 - val_acc: 0.9995
Epoch 18/20
8000/8000 [==============================] - 0s 22us/step - loss: 0.0314 - acc: 1.0000 - val_loss: 0.0302 - val_acc: 0.9995
Epoch 19/20
8000/8000 [==============================] - 0s 22us/step - loss: 0.0276 - acc: 1.0000 - val_loss: 0.0266 - val_acc: 0.9995
Epoch 20/20
8000/8000 [==============================] - 0s 21us/step - loss: 0.0244 - acc: 1.0000 - val_loss: 0.0235 - val_acc: 1.0000
----- activations -----
(1, 4)
attention = [0.05938202 0.7233456 0.1254946 0.09177781]05-11 20:54