Question about TFRecord data shuffle in DALI.

Question about TFRecord data shuffle in DALI

Hello~, I have some question about TFRecord data's random_shuffle in DALI.

For example, I have a dataset contains 8k images, when I make it a TFRecord data, it is spilt into 8 files like dataset.tfrecord-00000-of-00008, dataset.tfrecord-00001-of-00008... dataset.tfrecord-00007-of-00008, each of them contains 1k images.
When I use fn.readers.tfrecord(random_shuffle=True), how does it realize shuffle?
Situation 1: The 8 files random_shuffle in its own part, which can be thought as 8 separately random_shuffle, in each they random shuffle its own 1k images.
Situation 2: The 8 files random_shuffle together. They random shuffle 8k images together.

The reason I ask this question is because when I am using DALI do my training. Traing with DALI processing the data get a lower metric than trainging whithout DALI. But if I put random_shuffle=False and make they load data as the same order, the metric of them are nearly same. So I wondered if DALI TFRecord's random_shuffle maybe the reason causing the lower metric?

Thanks for your reading, it will help me a lot.

Check for duplicates

• I have searched the open bugs/issues and have found no duplicates for this bug report