Raise ValueError when number of dims evaluate to zero

[sampathweb]

When TransformerEncoder and TransformerDecoder is used in a Model with insufficient embedding_size for EmbeddingLayer, it can create divide by zero error.

For example -

model = Sequential(layers=[
    Embedding(50, embedding_size=2, input_length=10, name='embedding'),
    TransformerEncoder(intermediate_dim=64, num_heads=8),
    Flatten(),
    Dense(16, activation="relu"),
    Dropout(0.2),
    Dense(1, activation="linear")
])

Fails with divide by zero error since embedding_size of 2 // num_heads results in zero. This PR will provide a more meaning ful message on why this happens.

[jbischof]

Thanks @sampathweb! Is there an issue for this?

[sampathweb]

Thanks @sampathweb! Is there an issue for this?

Yes, a user who tried to do this and this code snippet is from that user's question. I will DM the issue.

[mattdangerw]

LGTM! Thanks!

[mattdangerw]

/gcbrun