如何构建堆叠式Keras LSTM？解决输入维度不兼容报错

问题背景 #

原非堆叠LSTM模型运行正常，代码如下：

# reshape input to be [samples, time steps, features]
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))

# create and fit the LSTM network
model = Sequential()
model.add(LSTM(4, input_shape=(1, LOOK_BACK)))

model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(trainX, trainY, epochs=EPOCHS, batch_size=1, verbose=2)

# make predictions
trainPredict = model.predict(trainX)
testPredict = model.predict(testX)

尝试构建堆叠LSTM时，使用如下代码出现报错：

# create and fit the LSTM network
model = Sequential()
batch_size = 1
model.add(LSTM(4, batch_input_shape=(batch_size, LOOK_BACK, 1), stateful=True, return_sequences=True))
model.add(LSTM(4, batch_input_shape=(batch_size, LOOK_BACK, 1), stateful=True))

报错信息：

ValueError: Input 0 of layer "lstm_6" is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: (None, 4)

错误原因 #

1. 第二层LSTM的batch_input_shape参数设置错误：第一层LSTM设置了return_sequences=True，输出形状为(batch_size, time_steps, units)即(1, LOOK_BACK, 4)，但第二层指定的batch_input_shape=(batch_size, LOOK_BACK, 1)和实际输入维度不匹配。
2. 堆叠LSTM时，只有第一层需要指定输入形状，后续层会自动适配前一层的输出形状，无需重复指定batch_input_shape或input_shape。

正确实现方式 #

方式1：无状态（stateful=False）堆叠LSTM（与原代码风格一致） #

这种方式无需固定batch_size，更适合常规场景：

# 保持原输入reshape逻辑不变
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))

model = Sequential()
# 第一层LSTM：设置return_sequences=True，输出3维张量供下一层LSTM使用
model.add(LSTM(4, input_shape=(1, LOOK_BACK), return_sequences=True))
# 第二层LSTM：无需指定输入形状，自动适配前一层输出
model.add(LSTM(4))
# 输出层
model.add(Dense(1))

model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(trainX, trainY, epochs=EPOCHS, batch_size=1, verbose=2)

trainPredict = model.predict(trainX)
testPredict = model.predict(testX)

方式2：有状态（stateful=True）堆叠LSTM #

若需要使用stateful模式，需固定batch_size，且注意输入形状的匹配：

batch_size = 1
# 输入reshape需确保样本数是batch_size的整数倍（stateful模式要求）
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))

model = Sequential()
# 第一层LSTM：指定batch_input_shape，return_sequences=True输出3维张量
model.add(LSTM(4, batch_input_shape=(batch_size, 1, LOOK_BACK), stateful=True, return_sequences=True))
# 第二层LSTM：无需指定batch_input_shape，自动适配前一层输出；若要指定，需匹配上一层输出形状(batch_size, 1, 4)
model.add(LSTM(4, stateful=True))
model.add(Dense(1))

model.compile(loss='mean_squared_error', optimizer='adam')
# stateful模式训练时，需手动重置状态
for epoch in range(EPOCHS):
    model.fit(trainX, trainY, epochs=1, batch_size=batch_size, verbose=2, shuffle=False)
    model.reset_states()

trainPredict = model.predict(trainX, batch_size=batch_size)
testPredict = model.predict(testX, batch_size=batch_size)

关键参数说明 #

• return_sequences=True：必须在堆叠的中间LSTM层设置，使层输出包含时间维度的3维张量(samples, time_steps, units)，供下一层LSTM处理；最后一层LSTM无需设置，直接输出2维张量(samples, units)给全连接层。
• stateful=True：启用状态保持时，需固定batch_size，训练时不能打乱数据（shuffle=False），且每个epoch结束后需手动调用model.reset_states()重置状态。

内容的提问来源于stack exchange，提问作者bbartling