Are images from different TF-records contained in different folders shuffled to form the training batches?

[rcruzgar]

Hi,

I am trying to understand how the object_detection API deals with the training TF-records.
As you know, TF-records can be provided specifying as many input_path items as wanted in the train_input_reader section of the pipeline.config file.

For example:

  tf_record_input_reader {
    input_path: "asia/train_*.record"
    input_path: "america/train_*.record"
    input_path: "europe/train_*.record"
  }

My question is: in which order are the TF-records used in the training, are they shuffled? I understand the images inside each TF-record are shuffled, aren't they? But are the images of different TF-records from different folders included on a same batch?

I am afraid that on each batch there are only images of one continent (as it is how the folders containing my TF-records are organized, i.e. Asia, America, Europe...).

Best regards,
Rubén.