1. simple RNN
下面创建一个简单的 2 层 RNN，每层有 100 个神经元，输出层是单个神经元的 dense 层：
model1 = keras.models.Sequential()
model1.add(keras.layers.SimpleRNN(100, return_sequences=True, input_shape=input_shape))
model1.add(keras.layers.SimpleRNN(100))
model1.add(keras.layers.Dense(1))
model1.compile(loss="mse", optimizer=keras.optimizers.SGD(lr=0.005), metrics=["mae"])
history1 = model1.fit(X_train_3D, y_train, epochs=200, batch_size=200,
validation_data=(X_valid_3D, y_valid))
2. Seq2Seq
建立一个 Seq2Seq 模型，和简单 RNN的区别是，第二个 RNN 层也用了 return sequences=True
model2 = keras.models.Sequential()
model2.add(keras.layers.SimpleRNN(100, return_sequences=True, input_shape=input_shape))
model2.add(keras.layers.SimpleRNN(100, return_sequences=True))
model2.add(keras.layers.Dense(1))
model2.compile(loss=huber_loss, optimizer=keras.optimizers.SGD(lr=0.01),
metrics=[mae_last_step])
history2 = model2.fit(X_train_3D, Y_train_3D, epochs=200, batch_size=200,
validation_data=(X_valid_3D, Y_valid_3D))
3. Seq2Seq LSTM
将普通 rnn 层换成 lstm 层
model3 = keras.models.Sequential()
model3.add(keras.layers.LSTM(100, return_sequences=True, input_shape=input_shape))
model3.add(keras.layers.LSTM(100, return_sequences=True))
model3.add(keras.layers.Dense(1))
model3.compile(loss=huber_loss, optimizer=keras.optimizers.SGD(lr=0.01),
metrics=[mae_last_step])
history3 = model3.fit(X_train_3D, Y_train_3D, epochs=200, batch_size=200,
validation_data=(X_valid_3D, Y_valid_3D),
callbacks=[keras.callbacks.ReduceLROnPlateau(verbose=1)])
4. 1D-ConvNets 预处理
在前面加上一个 1D-ConvNets 预处理
model4 = keras.models.Sequential()
model4.add(keras.layers.Conv1D(32, kernel_size=5, input_shape=input_shape))
model4.add(keras.layers.MaxPool1D(pool_size=5, strides=2))
model4.add(keras.layers.LSTM(32, return_sequences=True))
model4.add(keras.layers.LSTM(32))
model4.add(keras.layers.Dense(1))
model4.compile(loss=huber_loss, optimizer=keras.optimizers.SGD(lr=0.005))
history4 = model4.fit(X_train_3D, y_train, epochs=200, batch_size=100,
validation_data=(X_valid_3D, y_valid))