拟合LSTM模型

Question

我正在尝试拟合 LSTM 模型，但它给了我一个形状错误。

我的数据集有 218 行和 16 个特征，包括目标特征。我拆分了数据，%80 用于训练，%20 用于测试，在编译模型和运行之后，我得到了这个错误：

InvalidArgumentError:    Specified a list with shape [160,1] from a tensor with shape [14,1]
     [[{{node TensorArrayUnstack/TensorListFromTensor}}]]
     [[functional_7/lstm_6/PartitionedCall]] [Op:__inference_train_function_21740]
Function call stack:
train_function -> train_function -> train_function

变量定义： batch_size = 160 epochs = 20 timesteps = 15

整形后的训练集和测试集如下： y_train = (174, 15, 1) y_train = (174, 1, 1) x_test = (44, 15, 1) y_test = (44, 1, 1)

我的模型：

当我拟合模型时，此代码出现问题：

Answer 1

两件事：如果模型的输入和输出应该具有相同的形状，则必须更改 y_train 的形状（检查模型摘要）。其次，样本的数量，在你的例子中是 174，应该可以被 batch_size 整除而没有余数。因此，您只能使用 1、2、3、6、29、58、87 或 174 作为批量大小。这是一个工作示例：

import tensorflow as tf

batch_size = 2 
epochs = 20 
timesteps = 15

inputs_1_mae = tf.keras.layers.Input(batch_shape=(batch_size, timesteps, 1)) 
lstm_1_mae = tf.keras.layers.LSTM(100, stateful = True, return_sequences = True)(inputs_1_mae) 
lstm_2_mae = tf.keras.layers.LSTM(100, stateful = True, return_sequences = True)(lstm_1_mae) 
output_1_mae = tf.keras.layers.Dense(units = 1)(lstm_2_mae) 
regressor_mae = tf.keras.Model(inputs= inputs_1_mae ,outputs = output_1_mae) 
regressor_mae.compile (optimizer = "adam", loss = "mae") 
regressor_mae.summary() 

x_train = tf.random.normal((174, 15, 1)) 
y_train = tf.random.normal((174, 15, 1))

regressor_mae.fit(x_train, y_train, batch_size = batch_size, epochs=2)

Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(2, 15, 1)]              0         
                                                                 
 lstm (LSTM)                 (2, 15, 100)              40800     
                                                                 
 lstm_1 (LSTM)               (2, 15, 100)              80400     
                                                                 
 dense (Dense)               (2, 15, 1)                101       
                                                                 
=================================================================
Total params: 121,301
Trainable params: 121,301
Non-trainable params: 0
_________________________________________________________________
Epoch 1/2
87/87 [==============================] - 4s 5ms/step - loss: 0.8092
Epoch 2/2
87/87 [==============================] - 0s 5ms/step - loss: 0.8089
<keras.callbacks.History at 0x7f5820061250>

更新 1：要绘制训练和测试数据的均方误差，请尝试这样的操作：

x_train = tf.random.normal((174, 15, 1)) 
y_train = tf.random.normal((174, 15, 1))

x_test = tf.random.normal((174, 15, 1)) 
y_test = tf.random.normal((174, 15, 1))

history = regressor_mae.fit(x_train, y_train, batch_size = batch_size, epochs=25, validation_data=(x_test, y_test))

plt.plot(history.history['mean_absolute_error']) 
plt.plot(history.history['val_mean_absolute_error']) 
plt.title('model mean absolute error') 
plt.ylabel('mean_absolute_error') 
plt.xlabel('epoch') 
plt.legend(['train', 'test'], loc='upper left') 
plt.savefig('accuracy.png') 
plt.show()

拟合LSTM模型

Fitting LSTM model

python

lstm

keras

tensorflow