Keras LSTM输入和输出维度问题

Question

我正在尝试创建用于多步骤预测的LSTM模型。现在我正在测试模型网络设置，但它发现该设置存在维度问题。

这是我的测试数据集：

length = 100
df = pd.DataFrame()
df['x1'] = [i/float(length) for i in range(length)]
df['x2'] = [i**2 for i in range(length)]
df['y'] = df['x1'] + df['x2'] 
x_value = df.drop(columns = 'y').values
y_value = df['y'].values.reshape(-1,1)

这是我的t窗口数据构建功能：

def build_data(x_value, y_value ,n_input, n_output):
X, Y = list(), list()
in_start = 0
data_len = len(x_value)
# step over the entire history one time step at a time
for _ in range(data_len):
# define the end of the input sequence
in_end = in_start + n_input
out_end = in_end + n_output
if out_end <= data_len:
x_input = x_value[in_start:in_end] # e.g. t0-t3
X.append(x_input)
y_output = y_value[in_end:out_end] # e.g. t4-t5
Y.append(y_output)
# move along one time step
in_start += 1
return np.array(X), np.array(Y)            
X, Y = build_data(x_value, y_value, 1, 2)

X和Y 的形状

X.shape
### (98, 1, 2)
Y.shape
### (98, 2, 1)

对于模型部件，

verbose, epochs, batch_size = 1, 20, 16
n_neurons =  100
n_inputs, n_features  = X.shape[1], X.shape[2]
n_outputs = Y.shape[1]
model = Sequential()
model.add(LSTM(n_neurons, input_shape = (n_inputs, n_features), return_sequences=True))
model.add(TimeDistributed(Dense(1)))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(X, Y, epochs=epochs, batch_size=batch_size, verbose=verbose)

出现错误：ValueError: Error when checking target: expected time_distributed_41 to have shape (1, 1) but got array with shape (2, 1)

如果使用X, Y = build_data(x_value, y_value, 2, 2)i.e. input window == output window，则会起作用。但我认为它不应该包含这种限制。

我该如何克服这个问题？即input window != output window的设置

或者我应该设置的任何层或设置？

Answer 1

在处理时间维度时遇到形状不匹配。。。当你试图预测时间维度为2的东西时，时间输入dim为1。所以你的人际网络中需要一些能够从1到2时间维度增加的东西。我使用了Upsampling1D层，在完整的示例下面

# create fake data
X = np.random.uniform(0,1, (98,1,2))
Y = np.random.uniform(0,1, (98,2,1))
verbose, epochs, batch_size = 1, 20, 16
n_neurons =  100
n_inputs, n_features  = X.shape[1], X.shape[2]
n_outputs = Y.shape[1]
model = Sequential()
model.add(LSTM(n_neurons, input_shape = (n_inputs, n_features), return_sequences=True))
model.add(UpSampling1D(n_outputs))
model.add(TimeDistributed(Dense(1)))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(X, Y, epochs=epochs, batch_size=batch_size, verbose=verbose)

其中输入时间dim>输出temporal dim时，可以使用Lambda或Pooling操作(如果维度匹配(。下面是Lambda 的示例

X = np.random.uniform(0,1, (98,3,2))
Y = np.random.uniform(0,1, (98,2,1))
verbose, epochs, batch_size = 1, 20, 16
n_neurons =  100
n_inputs, n_features  = X.shape[1], X.shape[2]
n_outputs = Y.shape[1]
model = Sequential()
model.add(LSTM(n_neurons, input_shape = (n_inputs, n_features), return_sequences=True))
model.add(Lambda(lambda x: x[:,-n_outputs:,:]))
model.add(TimeDistributed(Dense(1)))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(X, Y, epochs=epochs, batch_size=batch_size, verbose=verbose)