Hardcoded bucket_step overrides configurable parameter

[iqddd]

Hey, just noticed that bucket_step is hardcoded to 64 here:

sd-scripts/train_network.py

Lines 127 to 130 in 0778dd9

	def assert_extra_args(self, args, train_dataset_group: Union[train_util.DatasetGroup, train_util.MinimalDataset], val_dataset_group: Optional[train_util.DatasetGroup]):
	train_dataset_group.verify_bucket_reso_steps(64)
	if val_dataset_group is not None:
	val_dataset_group.verify_bucket_reso_steps(64)

But there's already a configurable parameter for this elsewhere in the code. This kinda defeats the purpose of having it configurable in the first place.

Personally, I find 32 way more practical than 64. Any particular reason for this change? Any chance we can bring back configurability?

[iqddd]

verify_bucket_reso_steps(32) is unreachable in FluxNetworkTrainer

In FluxNetworkTrainer.assert_extra_args, we call super().assert_extra_args(args, train_dataset_group, val_dataset_group) first:

class FluxNetworkTrainer(train_network.NetworkTrainer):
# ------
    def assert_extra_args(self, args, train_dataset_group: Union[train_util.DatasetGroup, train_util.MinimalDataset], val_dataset_group: Optional[train_util.DatasetGroup]):
        # ↓↓↓↓↓↓↓↓↓↓↓↓↓↓
        super().assert_extra_args(args, train_dataset_group, val_dataset_group)
# ------
        train_dataset_group.verify_bucket_reso_steps(32)  # TODO check this

However, super().assert_extra_args (from NetworkTrainer) already calls:

class NetworkTrainer:
# ------
    def assert_extra_args(self, args, train_dataset_group: Union[train_util.DatasetGroup, train_util.MinimalDataset], val_dataset_group: Optional[train_util.DatasetGroup]):
        train_dataset_group.verify_bucket_reso_steps(64)
        if val_dataset_group is not None:
            val_dataset_group.verify_bucket_reso_steps(64)

Since verify_bucket_reso_steps(64) asserts if the value is lower than 64, the execution will never reach verify_bucket_reso_steps(32), making it effectively dead code.

Is this intentional, or should verify_bucket_reso_steps(32) be called before super().assert_extra_args?

[PercyDan54]

verify_bucket_reso_steps(32) is unreachable in FluxNetworkTrainer

It is also unreachable in SdxlNetworkTrainer and Sd3NetworkTrainer.

And it seems a regression from
fe2aa32#diff-62cf7de156b588b9acd7af26941d7bb189368221946c8b5e63f69df5cda56f39

[kohya-ss]

Thank you for this. I will check it sooner.

[rockerBOO]

https://github.com/black-forest-labs/flux/blob/main/demo_st_fill.py#L84-L105 confirming flux suggests possibly 32.

SD1.5 wants 64 (train_network.py version).