在 recurrentshop 中使用 return_sequences

Question

我正在尝试使用 recurrentshop 在 Keras 中编写自定义循环层。输入是一个长度为 "timesteps" 的序列，输出应该是所有时间步输出的序列，因此我使用 return_sequences=True.

循环模型有两个输入，其中一个是从最后一个时间步开始循环的。我正在使用类似于 reccurentshop example 的这一层，但我不断收到错误提示 "You must feed a value for placeholder tensor 'input_2' with dtype float"。我究竟做错了什么？完整代码如下：

import keras.backend as K
from recurrentshop import RecurrentModel
import numpy as np
from keras.models import Model
from keras.layers import Dense, Reshape, Conv1D, Input, Lambda, concatenate
from keras.optimizers import Adam

# parameters:
timesteps = 35
output_dim = 315
input_dim = 10
batch_size = 100

# recurrent layer definition:
def myRNN(input_dim,output_dim):
    inp = Input((input_dim,))
    h_tm1 = Input((output_dim,))
    modified_h = Lambda(lambda x: x * K.sum(K.square(inp)))(h_tm1)
    modified_inp = Dense(output_dim, use_bias=False, activation='tanh')(inp)
    modified_inp = Reshape((output_dim,1))(modified_inp)
    modified_inp = Conv1D(128, 7, padding='same', activation='tanh', use_bias=False)(modified_inp)
    modified_inp = Lambda(lambda x: K.sum(x, axis=-1))(modified_inp)
    hid = concatenate([modified_h, modified_inp], axis=-1)
    h_t = Dense(output_dim, use_bias=False, activation='tanh')(hid)  
    return RecurrentModel(input=inp, output=h_t, initial_states=h_tm1, final_states=h_t,
                          return_sequences=True, state_initializer=['zeros'])

# building the model:
inp = Input((timesteps, input_dim))
temp = myRNN(input_dim,output_dim)(inp)
out = Reshape((timesteps*output_dim,1))(temp)
model = Model(inputs=inp, outputs=out)
model.compile(loss='mse', optimizer='adam')

# testing the model:
inp = np.random.rand(batch_size ,timesteps ,input_dim)
prediction = model.predict(inp)

Answer 1

原来问题出在 Lambda 层：

modified_h = Lambda(lambda x: x * K.sum(K.square(inp)))(h_tm1)

通过定义函数解决了：

def factored_h(arg):
    norm = K.sum(K.square(arg[0]))
    return arg[1]*norm

并将 Lambda 层改写如下：

modified_h =Lambda(factored_h)([inp, h_tm1])

在 recurrentshop 中使用 return_sequences

using return_sequences in recurrentshop

keras

recurrent-neural-network