tensorflow time series mode training : Only one input size may be -1, not both 0 and 1

Question

I was trying to do a tutorial on time series model with tensorflow and I got an error regarding reshaping presumably coming from a reshape layer.

train_dataset = tf.keras.utils.timeseries_dataset_from_array(
    train_norm,
    targets=temperature_train[delay:],
    sampling_rate=sampling_rate,
    sequence_length=sequence_length,
    shuffle=True,
    batch_size=batch_size,
)

valid_dataset = tf.keras.utils.timeseries_dataset_from_array(
    valid_norm,
    targets=temperature_valid[delay:],
    sampling_rate=sampling_rate,
    sequence_length=sequence_length,
    shuffle=True,
    batch_size=batch_size,
)

for batch in train_dataset.take(1):
    inputs, label = batch
    print(inputs.shape)
    print(label.shape)
    break

The print results are

(64, 120, 14)
(64,)

The following assertion block raised no error for train_dataset and valid_dataset:

for batch in train_dataset:
    inputs, label = batch
    assert inputs.shape[1:] == (120, 14), f"{inputs.shape[1:]}"

When I build a model and start training

inputs = tf.keras.Input(shape=(inputs.shape[1:]))
x = tf.keras.layers.Flatten()(inputs)
x = tf.keras.layers.Dense(16, activation="relu")(x)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
model.compile("adam", loss="mse", metrics=["mae"])
callbacks = [
tf.keras.callbacks.ModelCheckpoint("jena_dense.keras", save_best_only=True)]

history = model.fit(train_dataset, validation_data = valid_dataset, epochs = 15, callbacks=callbacks )

It throws a following error

InvalidArgumentError: Graph execution error:

Detected at node functional_10_1/flatten_13_1/Reshape defined at (most recent call last)

...
...

Only one input size may be -1, not both 0 and 1
     [[{{node functional_10_1/flatten_13_1/Reshape}}]] [Op:__inference_multi_step_on_iterator_40343]

The input shape for the flatten layer shouldn't be unambiguous confirming from the assertion test. I don't know where it comes from. Could someone help me?

Sagar · Accepted Answer · 2026-04-21 07:36:08Z

0

It seems like your valid_dataset is actually completely empty because your validation array is too short to form a sequence after the delay slice. The assert block missed this because iterating over an empty dataset simply exits the loop immediately without running the assertion. Try to ensure len(temperature_valid[delay:]) >= sequence_length and the flatten error will disappear. Please let me know if it works for you. Thanks!

answered Apr 21 at 7:36

Sagar

4613 silver badges10 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

ComplexNumber Apr 22 at 15:10

Removing validation data in model.fit method gives the same result. history = model.fit( train_dataset, epochs=15 ) So it is not the case. And valid dataset is not at all empty after shortening with delay.

Sagar Apr 23 at 6:52

Did you try bypassing Flatten() entirely by replacing it with tf.keras.layers.Reshape((1680,))? Sometimes what happens is timeseries_dataset_from_array loses the static sequence length during graph execution which causes Flatten to crash. Hardcoding the math (120 * 14 = 1680) usually bypasses this.

ComplexNumber 2 days ago

Yes, reshaping works.. And this sounds like a bug

Frightera · Accepted Answer · 2026-04-25 14:35:48Z

0

Try Replacing

inputs = tf.keras.Input(shape=(inputs.shape\[1:\]))

with:

inputs = tf.keras.Input(shape=(120, 14))

[EDIT]

train_dataset = tf.keras.utils.timeseries_dataset_from_array(
    train_norm[:-delay],
    targets=temperature_train[delay:],
    sampling_rate=sampling_rate,
    sequence_length=sequence_length,
    shuffle=True,
    batch_size=batch_size,
)

valid_dataset = tf.keras.utils.timeseries_dataset_from_array(
    valid_norm[:-delay],
    targets=temperature_valid[delay:],
    sampling_rate=sampling_rate,
    sequence_length=sequence_length,
    shuffle=True,
    batch_size=batch_size,
)

edited Apr 25 at 14:35

Frightera

5,1542 gold badges15 silver badges31 bronze badges

answered Apr 17 at 7:13

Sharmistha

316 bronze badges

3 Comments

ComplexNumber Apr 17 at 7:42

Yes, as my comment on the other answer says, I tried this and it gives the same result.

Sharmistha Apr 17 at 9:40

You shifted targets forward but didn’t shorten inputs, so the dataset produces mismatched sample counts, which breaks batching and triggers the reshape error. I've added a potential fix to the original rsponse

ComplexNumber Apr 17 at 15:29

Actually, the implementation of the method takes care of it even if it goes without the shortening. Putting the full dataset or the shortened one does not make any difference. I tried it and it shows the same thing.

Collectives™ on Stack Overflow

tensorflow time series mode training : Only one input size may be -1, not both 0 and 1

2 Answers 2

3 Comments

3 Comments

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

3 Comments

3 Comments

Related