I finnaly find the problem.

tensorflow dataset are a kind of data generator so we don't need to use a data generator to chunck dataset and pass it to tensorflow dataset.

Use .batch() to generated "chunck" of data to read by itteration.