def neural_Conv1D(input_shape,
net_conv_num=[64, 64],
kernel_size=[5, 5],
pooling=True,
pooling_size=[5, 5],
net_dense_shape=[128, 64, 2],
optimizer_name='Adagrad',
lr=0.001):
'''
:param input_shape: 样本数据格式
:param net_shape: 神经网络格式
:param optimizer_name: 优化器
:param lr: 学习率
:param return: 返回神经网络模型
'''
model = Sequential()
# 增加Conv1D层
for n in range(len(net_conv_num)):
model.add(InputLayer(input_shape=input_shape))
model.add(
Conv1D(
filters=net_conv_num[n], # 卷积核数量
kernel_size=kernel_size[n], # 卷积核尺寸，或者[3]
strides=1,
padding='same',
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
name='Conv1D_' + str(n)))
if pooling == True:
model.add((
MaxPooling1D(
pool_size=pooling_size[n], # 卷积核尺寸，或者[3]
strides=2,
padding='same',
name='MaxPooling1D_' + str(n))))
# 增加展平层
model.add(Flatten())
# 增加全连接隐藏层
for n, units in enumerate(net_dense_shape[0:-1]):
model.add(
Dense(units=units,
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
name='Dense_' + str(n)))
# 增加最后的softmax层
model.add(
Dense(units=net_dense_shape[-1],
activation='softmax',
kernel_initializer=initializers.normal(stddev=0.1),
name='softmax'))
optimizer = optimizers(name=optimizer_name, lr=lr)
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def add_Conv1D(filters=16, # 卷积核数量
kernel_size=3, # 卷积核尺寸，或者[3]
strides=1,
padding='same',
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
**param):
model.add(Conv1D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
activation=activation,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
**param))
def add_Dense(units=16,
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
**param):
model.add(Dense(units=units,
activation=activation,
kernel_initializer=kernel_initializer,
**param))
def cnn1d(input_dim,
input_length=100,
output_dim=50,
label_n=6,
loss='categorical_crossentropy'):
'''
:param input_dim: 字典长度，即onehot的长度
:param input_length: 文本长度
:param output_dim: 词向量长度
:return:
'''
model = Sequential()
model.add(
Embedding(input_dim=input_dim + 1,
input_length=input_length,
output_dim=output_dim,
mask_zero=0))
model.add(
Conv1D(
filters=64, # 卷积核数量
kernel_size=5, # 卷积核尺寸，或者[3]
strides=1,
padding='same',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=2, padding='valid'))
model.add(
Conv1D(
filters=64, # 卷积核数量
kernel_size=5, # 卷积核尺寸，或者[3]
strides=1,
padding='same',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=2, padding='valid'))
model.add(Flatten(name='Flatten'))
model.add(Dense(units=64, activation='relu'))
model.add(Dropout(0.25))
model.add(Dense(units=label_n, activation='sigmoid'))
optimizer = Adagrad(lr=0.01)
model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])
return model
def neural_LSTM(input_shape,
net_shape=[64, 64, 128, 2],
optimizer_name='Adagrad',
lr=0.001):
'''
:param input_shape: 样本数据格式
:param net_shape: 神经网络格式
:param optimizer_name: 优化器
:param lr: 学习率
:param return: 返回神经网络模型
'''
model = Sequential()
# 识别之前的'截断/填充',跳过填充
model.add(Masking(mask_value=0, input_shape=input_shape))
#增加LSTM层
model.add(
LSTM(units=net_shape[0],
activation='tanh',
recurrent_activation='hard_sigmoid',
implementation=1,
dropout=0.2,
kernel_initializer=initializers.normal(stddev=0.1),
name='LSTM'))
#增加全连接隐藏层
for n, units in enumerate(net_shape[1:-1]):
model.add(
Dense(units=units,
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
name='Dense' + str(n)))
#增加最后的softmax层
model.add(
Dense(units=net_shape[-1],
activation='softmax',
kernel_initializer=initializers.normal(stddev=0.1),
name='softmax'))
optimizer = optimizers(name=optimizer_name, lr=lr)
model.compile(optimizer=optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def CuDNNGRU_drop(input_dim=100, input_length=100, output_dim=200, label_n=6):
'''
:param input_dim: 字典长度，即onehot的长度
:param input_length: 文本长度
:param output_dim: 词向量长度
:return:
'''
# input_dim = 100
# input_length = 100
# output_dim = 200
# label_n = 6
model = Sequential()
model.add(
Embedding(input_dim=input_dim + 1,
input_length=input_length,
output_dim=output_dim,
mask_zero=0))
model.add(CuDNNGRU(units=32, return_sequences=True))
model.add(
Conv1D(
filters=64, # 卷积核数量
kernel_size=5, # 卷积核尺寸，或者[3]
strides=1,
padding='same',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
activation='relu'))
model.add(MaxPooling1D(pool_size=2, strides=2, padding='valid'))
model.add(Flatten())
model.add(Dense(units=128, activation='tanh'))
model.add(Dropout(0.20))
model.add(Dense(units=label_n, activation='sigmoid'))
optimizer = Adagrad(lr=0.001)
model.compile(optimizer=optimizer,
loss=multiply_loss,
metrics=['accuracy'])
return model
def neural_bulit(net_shape,
optimizer_name='Adagrad',
lr=0.001,
loss='categorical_crossentropy'):
'''
:param net_shape: 神经网络格式
:param optimizer_name: 优化器
:param lr: 学习率
:param loss: 损失函数
:param return: 返回神经网络模型
'''
model = Sequential()
for n in range(len(net_shape)):
if net_shape[n]['name'] == 'InputLayer':
model.add(
InputLayer(input_shape=net_shape[n]['input_shape'],
name='num_' + str(n) + '_InputLayer'))
elif net_shape[n]['name'] == 'Dropout':
if 'rate' not in net_shape[n]:
net_shape[n].update({'rate': 0.2})
model.add(
Dropout(rate=net_shape[n]['rate'],
name='num_' + str(n) + '_Dropout'))
elif net_shape[n]['name'] == 'Masking':
model.add(Masking(mask_value=0))
elif net_shape[n]['name'] == 'LSTM':
if 'units' not in net_shape[n]:
net_shape[n].update({'units': 16})
if 'activation' not in net_shape[n]:
net_shape[n].update({'activation': 'tanh'})
if 'recurrent_activation' not in net_shape[n]:
net_shape[n].update({'recurrent_activation': 'hard_sigmoid'})
if 'dropout' not in net_shape[n]:
net_shape[n].update({'dropout': 0.})
if 'recurrent_dropout' not in net_shape[n]:
net_shape[n].update({'recurrent_dropout': 0.})
model.add(
LSTM(units=net_shape[n]['units'],
activation=net_shape[n]['activation'],
recurrent_activation=net_shape[n]['recurrent_activation'],
implementation=1,
dropout=net_shape[n]['dropout'],
recurrent_dropout=net_shape[n]['recurrent_dropout'],
name='num_' + str(n) + '_LSTM'))
elif net_shape[n]['name'] == 'Conv1D':
if 'filters' not in net_shape[n]:
net_shape[n].update({'filters': 16})
if 'kernel_size' not in net_shape[n]:
net_shape[n].update({'kernel_size': 3})
if 'strides' not in net_shape[n]:
net_shape[n].update({'strides': 1})
if 'padding' not in net_shape[n]:
net_shape[n].update({'padding': 'same'})
model.add(
Conv1D(
filters=net_shape[n]['filters'], # 卷积核数量
kernel_size=net_shape[n]['kernel_size'], # 卷积核尺寸，或者[3]
strides=net_shape[n]['strides'],
padding=net_shape[n]['padding'],
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
bias_initializer=initializers.normal(stddev=0.1),
name='num_' + str(n) + '_Conv1D'))
elif net_shape[n]['name'] == 'MaxPooling1D':
if 'pool_size' not in net_shape[n]:
net_shape[n].update({'pool_size': 3})
if 'strides' not in net_shape[n]:
net_shape[n].update({'strides': 1})
if 'padding' not in net_shape[n]:
net_shape[n].update({'padding': 'same'})
model.add(
MaxPooling1D(
pool_size=net_shape[n]['pool_size'], # 卷积核尺寸，或者[3]
strides=net_shape[n]['strides'],
padding=net_shape[n]['padding'],
name='num_' + str(n) + '_MaxPooling1D'))
elif net_shape[n]['name'] == 'Flatten':
model.add(Flatten())
elif net_shape[n]['name'] == 'Dense':
if 'units' not in net_shape[n]:
net_shape[n].update({'units': 16})
model.add(
Dense(units=net_shape[n]['units'],
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
name='num_' + str(n) + '_Dense'))
elif net_shape[n]['name'] == 'softmax':
if 'units' not in net_shape[n]:
net_shape[n].update({'units': 16})
model.add(
Dense(units=net_shape[n]['units'],
activation='softmax',
kernel_initializer=initializers.normal(stddev=0.1),
name='num_' + str(n) + '_softmax'))
optimizer = optimizers(name=optimizer_name, lr=lr)
model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])
return model
label_transform = np.array(pd.get_dummies(total.loc[:, 'label']))
print(data_transform.shape)
# 拆分为训练集和测试集
train_data, test_data, train_label, test_label = train_test_split(data_transform,
label_transform,
test_size=0.33,
random_state=42)
model = Sequential()
# 识别之前的'截断/填充',跳过填充
model.add(Masking(mask_value=0, input_shape=data_transform.shape[-2:]))
model.add(LSTM(units=64,
activation='relu',
implementation=1,
dropout=0.2,
kernel_initializer=initializers.normal(stddev=0.1),
name='LSTM'))
# model.add(LSTM(units=64,
# activation='relu',
# # dropout=0.01,
# implementation=1,
# dropout=0.2,
# name='LSTM'))
model.add(Dense(units=64,
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
name='Dense1'))
model.add(Dense(units=128,
activation='relu',
kernel_initializer=initializers.normal(stddev=0.1),
name='Dense2'))