python - Interleave TFRecordDatasets of different lengths

Question

I'd like to interleave TFRecordDatasets of different sizes after having processed them indipendently.

In other words, I need the output to be equal to the concatenatation of all parts. While parts may be shuffled, I need to preserve the order within each part. For instance, if parts are of the form abc, def, ghi, the ouput should be abcdefghi (or any combination like abcghidef that still preserves the internal order of each part).

I played a bit with the cycle_length and block_length options of the interleave API but I'm confused about how to use them with datasets of different sizes.

import tensorflow as tf

def _interleave_data(filename):
    ds_part = tf.data.TFRecordDataset(filename).map(_parse_features)
    ds_part = ds_part.map(_do_some_processing) 
    return ds_part

dataset = tf.data.Dataset.list_files(filenames)
ds = dataset.interleave(_interleave_data)

Is there any elegant way to do it? Thanks!

Answer 1

等待大神答复