Question: how best to set the epoch dependent parameters

[nikky4D]

For a new max epoch run, is there any recommendation for setting epoch related values like: flat_epoch, no_aug_epoch, policy epoch, mixup epoch and stop epoch?

## Our LR-Scheduler
flat_epoch: 29 # 29    # 4 + epoch // 2, e.g., 40 = 4 + 72 / 2
no_aug_epoch: 8 #25 # 8

train_dataloader: 
  dataset: 
    transforms:
      policy:
        epoch: [4, 29, 50] #[15, 76, 142] #[4, 29, 50]   # list 

  collate_fn:
    mixup_epochs: [4, 29] #[15, 76] #[4, 29]
    stop_epoch: 50

[ShihuaHuang95]

Thank you very much for your interest and attention to our work. If you like what we’ve done, please consider giving this repo a star.

The issue you mentioned is actually quite simple. Assuming the total training epochs are 50:

The first 4 epochs are for data augmentation warmup.
Half of the epochs are used for dense one-to-one (o2o), i.e., 50/2 = 25.
The flat epoch is the sum of data augmentation warmup and dense o2o, i.e., 25 + 4 = 29.
In D-FINE, 8 epochs are used for searching EMA, so we follow this setting. As a result, no_aug_epoch = 8, and stop_epoch = 50.
The policy is epochs@[4, 29, 50].
I hope this helps you better understand our work!