Question

我正在尝试复制和改编一些LSTM代码。

在模型model.fit的合适部分，标题中提到了错误。

这是错误的整个回溯：

Traceback (most recent call last):

  File "<ipython-input-61-fc0772e5553b>", line 4, in <module>
    show_inner=False)])

  File "C:\Users\fassi\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\engine\training.py", line 952, in fit
    batch_size=batch_size)

  File "C:\Users\fassi\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\engine\training.py", line 751, in _standardize_user_data
    exception_prefix='input')

  File "C:\Users\fassi\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\engine\training_utils.py", line 138, in standardize_input_data
    str(data_shape))

ValueError: Error when checking input: expected lstm_10_input to have shape (679, 1) but got array with shape (1, 1)

这是命令行：

model.fit(X_train, y_train, batch_size=batch_size, nb_epoch=epochs, 
          validation_split=0.05, verbose=verbose, callbacks=
          [lstm_progress(), TQDMCallback(outer_description="%-16s"%"Training LSTM",
                        show_inner=False)])

我已经捏过X_train的形状：（679,1,1）和y_train：（（679,1）的形状，然后确定，我重塑了{{1 }}至（679,1）：

X_train

，并且出现以下错误： X_train = X_train.reshape(-1,1)

我也尝试这样做：ValueError: Error when checking input: expected lstm_10_input to have 3 dimensions, but got array with shape (679, 1)，以便具有这样的形状X_train = X_train.reshape(1,-1,1)，但出现此错误：

(1, 679,1)

最小的可复制代码（如果需要）：

ValueError: Error when checking target: expected activation_2 to have 3 dimensions, but got array with shape (679, 1)

预先感谢

NB。我只在{min {1}}中输入了一个示例代码。如果您在错误中输入了其他数字而不是679，那是很正常的。

Answer 1

由于您正在使用batch_size = batch_size调用.fit，因此该模型期望的数据形式为（batch_size，679,1）

Answer 2

我将更改2件事： -LSTM input_shape =（1,1） -LSTM return_sequences = False

使用示例代码：重塑X_train.shape =（6,1,1）=（样本，时间步长，特征）后
y_train.shape =（6,1）

因此LSTM input_shape =（1,1）而不是（6,1）。含义input_shape =（时间步长，特征）

此外，LSTM具有return_sequences = True，因此它返回3个暗淡数据，而不是Dense层的2个暗淡数据。导致另一个错误：

检查目标时出错：预期的activation_2为3 尺寸，但数组的形状为（6，1）

表示对于return_sequences = True，在这种情况下，y_train的输出应为3 dim而不是2：（6,1）如果设置return_sequences = False，则可以。

使用示例代码：

import pandas as pd
import keras
from keras.models import Sequential
from keras.layers import LSTM, Dropout, Dense, Activation


from time import daylight
import datetime
import numpy as np

O = [0.701733664614, 0.699495411782, 0.572129320819, 0.613315597684, 0.58079660603, 0.596638918579, 0.48453382119]
Ab = [datetime.datetime(2018, 12, 11, 14, 0), datetime.datetime(2018, 12, 21, 10, 0), datetime.datetime(2018, 12, 21, 14, 0), datetime.datetime(2019, 1, 1, 10, 0), datetime.datetime(2019, 1, 1, 14, 0), datetime.datetime(2019, 1, 11, 10, 0), datetime.datetime(2019, 1, 11, 14, 0)]

data = pd.DataFrame(np.column_stack([O,Ab]),
             columns=['ndvi', 'datetime'])

decade_end = len(data)-1
decade_begin = int(decade_end - len(data)*90/100) 

data2 = data.loc[decade_begin:decade_end]


def load_data(time_series, train_perc, normalise_window, 
              decade_begin, decade_end, points=None):

    data2 = time_series.loc[decade_begin:decade_end]
    print("data2: ",data2)

    scaler = None 

    train = data2['ndvi'].values[0:-1]#.reshape(-1,1)
    print("train: ",train)
    shift_steps = 1
    train_targets = data2['ndvi'].shift(-shift_steps).values[0:-1].reshape(-1,1)
    train_reshaped = train.reshape(len(train), 1, 1)
    x_train, y_train = train_reshaped, train_targets
    x_test = data2['ndvi'].values[:-1].reshape(-1,1)
    y_test = data2['ndvi'].values[-1:].reshape(-1,1)


    return [x_train, y_train, scaler, x_test, y_test] 

def build(layers):
    model = Sequential()
    print("layers: ", layers)
    model.add(LSTM(
        input_shape=(1, 1),
        output_dim=int(layers[0]),
        return_sequences=False))
    model.add(Dropout(0.2))
    model.add(Dense(
        output_dim=layers[0]))
    model.add(Activation("linear"))

    model.compile(loss="mse", optimizer="rmsprop")
    return model


X_train, y_train, scaler = load_data(data, 1, False, 
                                         decade_begin, decade_end, points=None)[0:3]

print(X_train)
seq_len_in  = len(X_train)
batch_size  = 1

nb_hidden   = int((2/3)*len(data))   


X_train.shape
y_train.shape

model = build([1, seq_len_in, nb_hidden])
model.summary()

X_train, y_train, scaler = load_data(data, 2, False, 
                                         decade_begin, decade_end, points=None)[0:3]
model.fit(X_train, y_train, batch_size=batch_size, nb_epoch=10, 
          validation_split=0.05, verbose=1)

ValueError：检查输入时出错：预期lstm_10_input具有形状（679，1），但数组的形状为（1，1）

2 个答案: