Working code broke after deploying to new installation. ValueError: When using `stateful=True` in a RNN, the batch size must be static. Found dynamic batch size: sequence.shape=(None, xx, xx) about tensorflow HOT 4 OPEN

Hi @summa-code ,

• Ensure that your input data has a static batch size. When using stateful=True, the batch size should not change between batches. You can achieve this by padding or truncating sequences to a fixed length.
• When defining your LSTM layer, set stateful=True and specify the batch_input_shape parameter with a fixed batch size and sequence length.
  example:

model.add(LSTM(units=64, batch_input_shape=(batch_size, sequence_length, input_dim), stateful=True))

• If you are using timeseries_dataset_from_array to prepare your data, make sure that all sequences have the same length. You can achieve this by setting the sequence_length parameter when creating the dataset.
  example:

dataset = timeseries_dataset_from_array(data, targets, sequence_length=desired_length, batch_size=batch_size)

Thank you!

from tensorflow.

So that was the working code, but after I updated to latest in fresh ubuntu install, "batch_input_shape" not a recognized attribute.
This is what I am getting

ValueError: Unrecognized keyword arguments passed to LSTM: {'batch_input_shape': (xxx, xxx, xxx)}

And yes, I am using sequence_length, please refer to timeseries example on this Tensorflow site.
https://www.tensorflow.org/tutorials/structured_data/time_series

from tensorflow.

Facing the exact same issue with tensorflow 2.16.1 and Python 3.9.16

from tensorflow.

Looks like either it is a bug or they got ride of the function, but either way, it is like a bug in how they configure the shape internally for statefulness. And also, there are no examples on how to use stateful LSTM in addition to lack of documentation on statefulness.

from tensorflow.