Large dataset streaming memory leak in TensorFlow

“partially decode, stream and filter big data with tensorflow_datasets (tfds) I have two issues (Note that this code is generated in google colab): Issue 1 I want to stream the droid dataset, which is almost 2TB big. I want to only use data which matches my filter conditions. For that I load the whole dataset and compute a generator yielding the next data sample, which matches the conditions. So that I don't need to load the whole data into RAM and filter on the fly. This is working for a test da”