通过自定义 LSTM 时的形状错误
Posted
技术标签:
【中文标题】通过自定义 LSTM 时的形状错误【英文标题】:Shape error when passed custom LSTM 【发布时间】:2018-12-28 14:51:23 【问题描述】:我一直在尝试自定义 LSTM 层以进一步改进。但是在我的自定义 LSTM 之后,池化层出现了一个看起来很正常的错误。
我的环境是:
赢10 keras 2.2.0 python 3.6回溯(最近一次通话最后一次): 文件“E:/PycharmProjects/dialogResearch/dialog/classifier.py”,第 60 行,在 模型 = build_model(word_dict, args.max_len, args.max_sents, args.embedding_dim) 文件“E:\PycharmProjects\dialogResearch\dialog\model\keras_himodel.py”, 第 177 行,在 build_model l_dense = TimeDistributed(Dense(200))(l_lstm) 文件“C:\ProgramData\Anaconda3\lib\site-packages\keras\engine\topology.py”, 第 592 行,在 调用 self.build(input_shapes[0]) 文件“C:\ProgramData\Anaconda3\lib\site-packages\keras\layers\wrappers.py”, 第 162 行,在构建中 断言 len(input_shape) >= 3 断言错误
我的自定义LSTM的代码是:
class CustomLSTM(Layer):
def __init__(self, output_dim, return_sequences, **kwargs):
self.init = initializers.get('normal')
# self.input_spec = [InputSpec(ndim=3)]
self.output_dim = output_dim
self.return_sequences = return_sequences
super(CustomLSTM, self).__init__(**kwargs)
def build(self, input_shape):
assert len(input_shape) == 3
self.original_shape = input_shape
self.Wi = self.add_weight('Wi', (input_shape[-1], self.output_dim), initializer=self.init, trainable=True)
self.Wf = self.add_weight('Wf', (input_shape[-1], self.output_dim), initializer=self.init, trainable=True)
self.Wo = self.add_weight('Wo', (input_shape[-1], self.output_dim), initializer=self.init, trainable=True)
self.Wu = self.add_weight('Wu', (input_shape[-1], self.output_dim), initializer=self.init, trainable=True)
self.Ui = self.add_weight('Ui', (self.output_dim, self.output_dim), initializer=self.init, trainable=True)
self.Uf = self.add_weight('Uf', (self.output_dim, self.output_dim), initializer=self.init, trainable=True)
self.Uo = self.add_weight('Uo', (self.output_dim, self.output_dim), initializer=self.init, trainable=True)
self.Uu = self.add_weight('Uu', (self.output_dim, self.output_dim), initializer=self.init, trainable=True)
self.bi = self.add_weight('bi', (self.output_dim,), initializer=self.init, trainable=True)
self.bf = self.add_weight('bf', (self.output_dim,), initializer=self.init, trainable=True)
self.bo = self.add_weight('bo', (self.output_dim,), initializer=self.init, trainable=True)
self.bu = self.add_weight('bu', (self.output_dim,), initializer=self.init, trainable=True)
super(CustomLSTM, self).build(input_shape)
def step_op(self, step_in, states):
i = K.softmax(K.dot(step_in, self.Wi) + K.dot(states[0], self.Ui) + self.bi)
f = K.softmax(K.dot(step_in, self.Wf) + K.dot(states[0], self.Uf) + self.bf)
o = K.softmax(K.dot(step_in, self.Wo) + K.dot(states[0], self.Uo) + self.bo)
u = K.tanh(K.dot(step_in, self.Wu) + K.dot(states[0], self.Uu) + self.bu)
c = i * u + f * states[1]
h = o * K.tanh(c)
return h, [h, c]
def call(self, x, mask=None):
init_states = [K.zeros((K.shape(x)[0], self.output_dim)),
K.zeros((K.shape(x)[0], self.output_dim))]
outputs = K.rnn(self.step_op, x, init_states)
if self.return_sequences:
return outputs[1]
else:
return outputs[0]
def compute_output_shape(self, input_shape):
return input_shape[0], input_shape[-1]
型号是:
def build_model(words, max_len, max_sents, embedding_dim):
sentence_input = Input(shape=(max_len,), dtype='int32')
embedding_layer = Embedding(len(words) + 1,
embedding_dim,
input_length=max_len,
trainable=True)
embedded_sequences = embedding_layer(sentence_input)
l_lstm = CustomLSTM(200, return_sequences=True)(embedded_sequences)
print(l_lstm.get_shape())
l_dense = TimeDistributed(Dense(200))(l_lstm)
l_att = AttLayer()(l_dense)
sentEncoder = Model(sentence_input, l_att)
review_input = Input(shape=(max_sents, max_len), dtype='int32')
review_encoder = TimeDistributed(sentEncoder)(review_input)
l_lstm_sent = CustomLSTM(200, return_sequences=True)(review_encoder)
l_dense_sent = TimeDistributed(Dense(200))(l_lstm_sent)
l_att_sent = AttLayer()(l_dense_sent)
preds = Dense(3, activation='softmax')(l_att_sent)
model = Model(review_input, preds)
optimizer = Nadam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=1e-08, schedule_decay=0.004)
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=[precision, recall, f1, 'acc'])
return model
感谢您的帮助。
【问题讨论】:
【参考方案1】:我认为发生错误是因为compute_output_shape
在return_sequences=True
返回的形状不正确。我会尝试以下方法:
def compute_output_shape(self, input_shape):
if self.return_sequences:
return input_shape
return (input_shape[0], input_shape[-1])
【讨论】:
你写的代码也让我了解这个函数是如何工作的以上是关于通过自定义 LSTM 时的形状错误的主要内容,如果未能解决你的问题,请参考以下文章
用户未经授权时的 Laravel 护照自定义错误消息和状态码