我正在尝试使用R中的Keras包为时间序列预测构建一维LSTM。我总共有280个观测值，并且正在前200个时间步上训练模型。
这是我的数据：
>data
[1] 104.674211 30.325553 25.725943 30.330470 33.919494 98.885751 31.036799 26.120383 29.929011 33.808716 102.614454 31.275918 28.122760
[14] 34.533204 38.933596 128.535598 31.998126 34.371124 35.299264 109.320850 37.147248 31.808292 36.188401 38.587492 110.945496 37.613322
[27] 31.879796 36.953585 39.512948 112.558327 35.770147 32.377192 37.855373 41.062163 115.672291 39.300869 34.376489 40.510942 44.743456
[40] 110.583600 45.455882 38.806954 46.087124 55.842239 153.782248 40.906253 46.261350 46.811348 125.209661 41.185677 36.597384 44.202294
[53] 47.642739 128.715730 41.212162 35.633601 47.998697 49.195762 128.057030 41.817710 36.760703 46.538546 47.615336 127.419624 44.014974
[66] 38.174168 46.570199 45.621642 5.754298 151.115306 40.753132 43.220820 45.130268 123.692596 41.815542 35.708891 44.339171 49.007662
[79] 127.696261 41.360583 36.686370 44.653017 47.199684 128.268945 43.121450 37.955169 47.769333 50.841333 164.244415 44.517725 44.149837
[92] 44.434503 131.707918 38.443468 33.104037 41.244507 45.168913 135.171285 38.445318 36.950023 42.376044 44.458236 131.206613 39.028980
[105] 35.517673 42.078143 46.161538 138.633018 40.927631 36.198797 46.439896 52.050011 158.697836 38.882071 40.011038 43.325122 131.811553
[118] 36.905217 32.293378 39.500356 45.379647 134.351935 38.506473 33.520961 40.085801 46.334602 141.613712 38.471661 34.943294 39.906911
[131] 43.418520 132.690750 37.797729 32.547215 41.195844 45.958671 161.540738 37.320867 40.446669 44.155767 138.347426 39.871351 33.207285
[144] 41.884156 53.102881 140.518116 41.748416 35.454985 45.002361 48.146979 147.215505 41.641961 37.037602 44.843474 48.614302 145.899917
[157] 39.998814 39.289558 47.533372 51.855770 155.500667 47.670483 46.645426 59.782910 71.429109 179.485628 47.311633 46.626551 141.468447
[170] 46.678339 43.436311 41.047591 113.229233 51.382804 31.984584 37.683744 41.678862 129.190309 36.947320 30.810330 39.766751 46.827086
[183] 134.329501 34.619556 31.565425 38.311338 41.986885 132.490021 34.646048 31.747129 40.542554 43.372881 129.704915 36.685621 30.791035
[196] 36.475763 43.158314 123.691776 41.041269 32.554182 38.578906 43.003428 6.936479 151.447335 35.098190 39.185031 41.954784 129.592065
[209] 37.503257 33.010408 41.009276 43.435023 127.724863 37.358387 33.076456 38.507582 44.386869 130.483615 37.264937 30.637003 41.278207
[222] 46.165747 132.370094 40.618600 34.049411 39.469918 42.530621 134.158590 41.086867 37.480499 58.861908 160.573323 40.978109 31.784450
[235] 40.325526 44.253421 132.244392 37.118805 30.989261 39.009905 44.869277 132.788785 36.757391 32.615994 40.015909 46.235674 134.601224
[248] 37.946823 33.413523 42.159784 45.313646 137.384025 39.622680 33.547060 42.216867 46.148574 135.930897 39.762899 34.289215 43.038315
[261] 48.735232 137.753104 42.082613 38.095095 48.786945 53.691134 171.482539 43.339101 44.989966 46.820114 135.725437 42.084594 35.534636
[274] 44.720649 49.923844 141.186341 41.887333 36.098428 45.116291 48.133853
拆分数据并建立模型：
train <- data[1:200]
test <- data[201:length(data)]
lstm_num_timesteps <- 1
X_train <- t(sapply(1:(length(train) - lstm_num_timesteps), function(x)
train[x:(x + lstm_num_timesteps - 1)]))
X_test <- t(sapply(1:(length(test) - lstm_num_timesteps), function(x) test[x:
(x + lstm_num_timesteps - 1)]))
y_train <- sapply((lstm_num_timesteps + 1):(length(train)), function(x) train[x])
y_test <- sapply((lstm_num_timesteps + 1):(length(test)), function(x) test[x])
X_train <- k_expand_dims(X_train, axis = 3)
# Building the model
model <- keras_model_sequential() %>%
layer_lstm(units = 5, batch_input_shape = c(200, 5, 1), stateful = TRUE, batch_size = 5, return_sequences = TRUE) %>%
layer_dense(units = 1)
history <- model %>% compile(
optimizer = 'rmsprop',
loss = 'categorical_crossentropy',
metrics = c('accuracy')
)
model %>% fit(X_train, y_train, batch_size = 10, epochs = 20, validation_data = list(X_test, y_test))
一切都很好，直到最终拟合线，出现以下错误：
Error in py_call_impl(callable, dots$args, dots$keywords) :
ValueError: Error when checking input: expected lstm_37_input to have 3 dimensions, but got array with shape (199, 5)
我尝试在第一个lstm层和最后一个密集层之间添加一个layer_flatten，但这也不起作用。不知道我是否在计算图层输入-高度赞赏任何建议！