Time Series data to fit for ConvLSTM-CodePudding

I used stock data with 4057 samples, made it into 28 time steps, with 25 features.

TrainX shape: (4057, 28, 25)

The Target consists of 5 categories of interger

[0,1,2,3,4]

and reshape into:

trainX_reshape= trainX.reshape(4057,1, 28,25,1)
testX_reshape= testX.reshape(1334,1, 28,25,1)

trying to fit the model:

seq =Sequential([
    ConvLSTM2D(filters=40, kernel_size=(3, 3),input_shape=(1, 28, 25, 1),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    Conv3D(filters=5, kernel_size=(3, 3, 3),activation='sigmoid',padding='same', data_format='channels_last')
])

compile with

seq.compile(loss='sparse_categorical_crossentropy', optimizer='rmsprop')

history = seq.fit(trainX_reshape, trainY, epochs=10,
               batch_size= 128, shuffle=False, verbose = 1,
               validation_data=(testX_reshape, testY),
               # validation_split=0.2)

and it gives ERROR:

InvalidArgumentError: Graph execution error:

How to fix it? Ive tried many methods, but had no clue.

the code and data are at: https://drive.google.com/drive/folders/1WDa_CUO1Mr7wZTqE3wHsR0Tp_3NRMcZ6?usp=sharing

works on colab

CodePudding user response：

Your model's output does not make any sense, if you are working with sparse integer labels. It is 5D and your labels are 2D (including batch size). Try:

seq =Sequential([
    ConvLSTM2D(filters=40, kernel_size=(3, 3),input_shape=(1, 28, 25, 1),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    ConvLSTM2D(filters=40, kernel_size=(3, 3),padding='same', return_sequences=True),
    BatchNormalization(),
    Conv3D(filters=5, kernel_size=(3, 3, 3),padding='same', data_format='channels_last'),
    GlobalMaxPooling3D(),
    Dense(5, activation='softmax')
])

Note, in your solution, the output of your model is reshaped to a 2D tensor to be compatible with sparse_categorical_crossentropy, but this reshaping is destroying the batch size. So you get the logits shape [89600,5] and labels shape [128], although it should be [128, 5] and [128]. If your model's output were, for example, [128, 1, 1, 1, 5], it would probably be possible.