rFID very high

[toyot-li]

Hi @yucornetto , I run the tokenizer training with

WANDB_MODE=offline WORKSPACE=/path/to/workspace accelerate launch      scripts/train_bar_fsq.py     config=configs/tokenizer/bar_fsq_16bits.yaml     dataset.params.train_shards_path_or_url=data/imagenet/imagenet1k-train-{0000..1023}.tar     dataset.params.eval_shards_path_or_url=data/imagenet/imagenet1k-validation-{00..63}.tar

The webdataset is directly from https://huggingface.co/datasets/timm/imagenet-1k-wds

but the eval metric is abnormal

{'InceptionScore': 17.84549007806725,
'LPIPS': 0.6580199621398378,
'MAE': 0.2317499674847123,
'MSE': 0.08625390160251671,
'PSNR': 11.101195318911644,
'SSIM': 0.19874112481006478,
'rFID': 155.83704336260826}

log0.txt

Training log is attached.

[yucornetto]

Hi, thanks for your interests in our work.

It looks like in the logging you only used one tar file for train & val respectively, please double check

train_shards_path_or_url: data/imagenet/imagenet1k-train-1023.tar
eval_shards_path_or_url: data/imagenet/imagenet1k-validation-63.tar

[toyot-li]

Hi, thanks for your interests in our work.

It looks like in the logging you only used one tar file for train & val respectively, please double check

train_shards_path_or_url: data/imagenet/imagenet1k-train-1023.tar
eval_shards_path_or_url: data/imagenet/imagenet1k-validation-63.tar

@yucornetto Thanks for your prompt response!

May I know how to appropriately configure the webdataset?

I use these args in the launching command, following README of this repo, as mentioned in the beginning.
dataset.params.train_shards_path_or_url=data/imagenet/imagenet1k-train-{0000..1023}.tar dataset.params.eval_shards_path_or_url=data/imagenet/imagenet1k-validation-{00..63}.tar

---

update: I write "dataset.params.train_shards_path_or_url=data/imagenet/imagenet1k-train-{0000..1023}.tar dataset.params.eval_shards_path_or_url=data/imagenet/imagenet1k-validation-{00..63}.tar" directly into configs/tokenizer/bar_fsq_16bits.yaml, then the log seems correct now.