PositionEmbedding Dimension Error

[Joe-Sedlacek]

Hi there,

Any help on this issue is greatly appreciated. I have been following the source code for keras_nlp.layers.PositionEmbedding() (tutorial found here https://keras.io/api/keras_nlp/modeling_layers/position_embedding/), but I am coming across an error which I am having trouble debugging. The error reads,

TypeError: Exception encountered when calling PositionEmbedding.call().

Dimension value must be integer or None or have an index method, got value '<attribute 'shape' of 'numpy.generic' objects>' with type '<class 'getset_descriptor'>'

Arguments received by PositionEmbedding.call():
• inputs=tf.Tensor(shape=(None, 256, 128), dtype=float32)
• start_index=0

The code that triggers this error is below:
inputs = layers.Input((config.MAX_LEN,), dtype="int64") word_embeddings = keras.layers.Embedding( config.VOCAB_SIZE, config.EMBED_DIM, name="word_embedding" )(inputs) position_embeddings = keras_nlp.layers.PositionEmbedding( sequence_length=config.MAX_LEN, )(word_embeddings) embeddings = word_embeddings + position_embeddings

config.MAX_LEN is an integer (I have checked its type), and the word_embeddings have shape (None, 256, 128), so I think they are getting created correctly?

Again, any help is much appreciated!