advise on stage 2

[rom1504]

• img2dataset
• audio2dataset
• video2dataset
• also have a all2dataset that does no re-encoding

important note is random shuffling allow to estimate size of whole collection quickly