meaning of length_bucket_options

[YandanYang]

Hi,

Could you please clarify the design rationale behind the length_bucket_options configuration? Specifically:

1. Why is the interval between consecutive values set to approximately 20 in the higher range?
2. What is the significance of 193 appearing in both lists, especially since it breaks the otherwise regular step pattern?

For reference, the current configuration is:

length_bucket_options = {
    1: [501, 481, 461, 441, 421, 401, 381, 361, 341, 321, 301, 281, 261, 241, 221, 193, 181, 161, 141, 121, 101, 81, 61, 41, 21],
    2: [193, 177, 161, 156, 145, 133, 129, 121, 113, 109, 97, 85, 81, 73, 65, 61, 49, 37, 25],
}

[SHYuanBest]

@YandanYang Thanks for your interest! Those numbers don't have any real meaning and were just chosen at random. The best practice to maximize data utilization is to match the bucket length with the training frame length, since our training frames are fixed at 16 + 2 + 1 + 9.