我是神经网络的新手,我想用它们与其他机器学习方法进行比较。我有一个大约两年的多元时间序列数据。我想根据其他变量使用 LSTM 预测接下来几天的“y”。我的数据最后一天是2020-07-31。
df.tail()
y holidays day_of_month day_of_week month quarter
Date
2020-07-27 32500 0 27 0 7 3
2020-07-28 33280 0 28 1 7 3
2020-07-29 31110 0 29 2 7 3
2020-07-30 37720 0 30 3 7 3
2020-07-31 32240 0 31 4 7 3
为了训练 LSTM 模型,我还将数据分为训练数据和测试数据。
from sklearn.model_selection import train_test_split
split_date = '2020-07-27' #to predict the next 4 days
df_train = df.loc[df.index <= split_date].copy()
df_test = df.loc[df.index > split_date].copy()
X1=df_train[['day_of_month','day_of_week','month','quarter','holidays']]
y1=df_train['y']
X2=df_test[['day_of_month','day_of_week','month','quarter','holidays']]
y2=df_test['y']
X_train, y_train =X1, y1
X_test, y_test = X2,y2
因为我正在使用 LSTM,所以需要一些缩放:
scaler = MinMaxScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)
现在,进入困难的部分:模型。
num_units=50
activation_function = 'sigmoid'
optimizer = 'adam'
loss_function = 'mean_squared_error'
batch_size = 10
num_epochs = 100
# Initialize the RNN
regressor = Sequential()
# Adding the input layer and the LSTM layer
regressor.add(LSTM(units = num_units, return_sequences=True ,activation = activation_function,
input_shape=(X_train.shape[1], 1)))
# Adding the output layer
regressor.add(Dense(units = 1))
# Compiling the RNN
regressor.compile(optimizer = optimizer, loss = loss_function)
# Using the training set to train the model
regressor.fit(X_train_scaled, y_train, batch_size = batch_size, epochs = num_epochs)
但是,我收到以下错误:
ValueError: Input 0 of layer sequential_11 is incompatible with the layer: expected ndim=3, found
ndim=2. Full shape received: [None, 5]
我不明白我们如何选择参数或输入的形状。我看过一些视频并阅读了一些 Github 页面,每个人似乎都以不同的方式运行 LSTM,这使得实现起来更加困难。之前的错误可能来自形状,但除此之外,其他一切都对吗?我该如何解决这个问题?谢谢
编辑:This 类似的问题并没有解决我的问题。我已经尝试了那里的解决方案
x_train = X_train_scaled.reshape(-1, 1, 5)
x_test = X_test_scaled.reshape(-1, 1, 5)
(我的 X_test 和 y_test 只有一列)。而且该解决方案似乎也不起作用。我现在收到此错误:
ValueError: Input 0 is incompatible with layer sequential_22: expected shape=
(None, None, 1), found shape=[None, 1, 5]
答案 0 :(得分:2)
输入:
问题是您的模型期望形状为 (batch, sequence, features)
的 3D 输入,但您的 X_train
实际上是一个数据框切片,因此是一个 2D 数组:
X1=df_train[['day_of_month','day_of_week','month','quarter','holidays']]
X_train, y_train =X1, y1
我假设你的列应该是你的特征,所以你通常会做的是你的 df 的“堆叠切片”,这样你 X_train
看起来像这样:
这是一个形状为 (15,5)
的虚拟二维数据集:
data = np.zeros((15,5))
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
您可以对其进行整形以添加批次维度,例如 (15,1,5)
:
data = data[:,np.newaxis,:]
array([[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]],
[[0., 0., 0., 0., 0.]]])
相同的数据,但以不同的方式呈现。现在在这个例子中,batch = 15
和 sequence = 1
,我不知道你的情况下的序列长度是多少,但它可以是任何东西。
模型:
现在在您的模型中,keras
input_shape
期望 (batch, sequence, features)
,当您通过此:
input_shape=(X_train.shape[1], 1)
这是您模型看到的:(None, Sequence = X_train.shape[1] , num_features = 1)
None
用于批次维度。我不认为这就是您在重塑后想要做的事情,您还应该更正 input_shape
以匹配新数组。
答案 1 :(得分:1)
这是您使用 LSTM 解决的多元回归问题。在进入代码之前,让我们实际看看它的含义
5
个功能 holidays, day_of_month, day_of_week,month,quarter
,持续 k
天y
天的 n
p
条记录没有可用的预测,其中 p
是序列长度。timeseries_dataset_from_array
方法创建窗口数据集。所以我们想要实现的图片如下所示:
对于每个 LSTM 单元展开,我们传入当天的 5 个特征,并在 m
时间展开,其中 m
是序列长度。我们正在预测最后一天的 y
。
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, models
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
# Model
regressor = models.Sequential()
regressor.add(layers.LSTM(5, return_sequences=True))
regressor.add(layers.Dense(1))
regressor.compile(optimizer='sgd', loss='mse')
# Dummy data
n = 10000
df = pd.DataFrame(
{
'y': np.arange(n),
'holidays': np.random.randn(n),
'day_of_month': np.random.randn(n),
'day_of_week': np.random.randn(n),
'month': np.random.randn(n),
'quarter': np.random.randn(n),
}
)
# Train test split
train_df, test_df = train_test_split(df)
print (train_df.shape, test_df.shape)\
# Create y to be predicted
# given last n days predict todays y
# train data
sequence_length = 3
y_pred = train_df['y'][sequence_length-1:].values
train_df = train_df[:-2]
train_df['y_pred'] = y_pred
# Validataion data
y_pred = test_df['y'][sequence_length-1:].values
test_df = test_df[:-2]
test_df['y_pred'] = y_pred
# Create window datagenerators
# Train data generator
train_X = train_df[['holidays','day_of_month','day_of_week','month','month']]
train_y = train_df['y_pred']
train_dataset = tf.keras.preprocessing.timeseries_dataset_from_array(
train_X, train_y, sequence_length=sequence_length, shuffle=True, batch_size=4)
# Validation data generator
test_X = test_df[['holidays','day_of_month','day_of_week','month','month']]
test_y = test_df['y_pred']
test_dataset = tf.keras.preprocessing.timeseries_dataset_from_array(
test_X, test_y, sequence_length=sequence_length, shuffle=True, batch_size=4)
# Finally fit the model
regressor.fit(train_dataset, validation_data=test_dataset, epochs=3)
输出:
(7500, 6) (2500, 6)
Epoch 1/3
1874/1874 [==============================] - 8s 3ms/step - loss: 9974697.3664 - val_loss: 8242597.5000
Epoch 2/3
1874/1874 [==============================] - 6s 3ms/step - loss: 8367530.7117 - val_loss: 8256667.0000
Epoch 3/3
1874/1874 [==============================] - 6s 3ms/step - loss: 8379048.3237 - val_loss: 8233981.5000
<tensorflow.python.keras.callbacks.History at 0x7f3e94bdd198>