CODEWITHSUNDEEP
pythontensorflowkerastime-serieslstm
Requin
Requin

Reputation: 577

Split tensorflow BatchDataset for LSTM with multiple inputs

I construct a LSTM model with two inputs: one for categorical variables, one for numerical variables:

model = Model(inputs = [cat_input, num_input], outputs = x, name = "LSTM")

The input data for the LSTM is generated by means of tensorflow.keras.utils.timeseries_dataset_from_array():

input_dataset = timeseries_dataset_from_array(
    df[["cat", "num1", "num2"]], df["target"], sequence_length=n_timesteps, sequence_stride=1, batch_size=20
)

When I directly feed input_dataset into the model, I get the following error: "ValueError: Layer "LSTM" expects 2 input(s), but it received 1 input tensors. Inputs received: [<tf.Tensor 'IteratorGetNext:0' shape=(None, None, 3) dtype=int64>]", because the model expects two inputs and not one.

I can achieve this (a bit ugly) like so:

input_dataset2 = input_dataset.map(lambda x, y: ((x[:,:,0:1], x[:,:,1:3]), y))
model.fit(
    input_dataset2, steps_per_epoch=20, epochs=50, verbose=0, shuffle=True
) # this now works

My question: The solution I found is not very elegant. Is this kind of split also possible with tf.split() or another function?

EDIT: When I try the following:

input_dataset.map(lambda x, y: ((split(value=x, num_or_size_splits=[1, 2], axis = -1)), y))

I get this error: "ValueError: Value [<tf.Tensor 'split:0' shape=(None, None, 1) dtype=int64>, <tf.Tensor 'split:1' shape=(None, None, 2) dtype=int64>] is not convertible to a tensor with dtype <dtype: 'int64'> and shape (2, None, None, None)."

Upvotes: 0

Views: 230

Answers (1)

Requin
Requin

Reputation: 577

I solved it by explicitly adding tuple() - the brackets are otherwise not recognized as tuple:

input_dataset = input_dataset.map(lambda x, y: 
    (
        tuple(
            tf.split(x, [1,2], -1)
        ),
        y
    )
)

Upvotes: 0

Related Questions