how to create dataset with label

[Johnson-yue]

If I want to finetune stylegan2-ada with conditional label how I prepare dataset ??

the readem had only for cifar10 dataset with conditional but how to use for my own dataset??

what is the format of data ?

[aiXander]

Here is my jupter notebook to create conditional datasets.

Note: this will not work out-of-the-box with StyleGAN's default dataloader, I'm using a tf_raw format, simply replace the dataset_tool.py file with this one and it should work.

Main benefit of the raw_tfrecords is that they are about ~10 times smaller than the original tfrecs which is insanely helpful if you ask me...

[Johnson-yue]

@tr1pzz thank you , I will try it ．

If I want to finetune stylegan2 checkpoint with conditional dataset , can I still use the same pretrained checkpoints with unconditional dataset？

[Johnson-yue]

@tr1pzz Hi, my data is gray image ,so it PIL.image.open(file) have a (256, 256) and
I replace the original dataset_tool.py with your dataset_tool.py and then use your jupyter notebook create a tfrecord successfully!

but when I try finetuning , has a error about dataset.load_dataset() function

****/training/dataset.py", line 219, in parse_tfrecord_np
    data = ex.features.feature["data"].bytes_list.value[0]    # pylint : disable=no-member original
IndexError: list index (0) out of range

then I debug and found :
ex.features.feature["data"].bytes_list.value[0] == None
but
ex.features.feature["img"].bytes_list.value[0] is good data
I can not load the tfrecord, why ?

original dataset_tool.py output is multi-resolution tfrreconrd file ,but my create is only one max-resolution tfrecord file is this a problem?

[aiXander]

Yeah, forgot about that, you also have to update a few minor things in dataset.py
You can take a look at the changes in this repo.
Let me know if you get stuck, it should be an easy fix to figure out!

[Johnson-yue]

@tr1pzz Hi ，I rewrite the data_tools.py reference create_cifar10 dataset ，and not change the dataset.py， Thank you for your help

[ivanlen]

Hey @Johnson-yue , were you able to train the conditional version from a pretrained unconditional one?

[ivanlen]

Hi @aiXander , can you share again the notebook to generate a conditioned GAN from an uncoditioned?
cheers!

[jazz-ai]

Here is my jupter notebook to create conditional datasets.

Note: this will not work out-of-the-box with StyleGAN's default dataloader, I'm using a tf_raw format, simply replace the dataset_tool.py file with this one and it should work.

Main benefit of the raw_tfrecords is that they are about ~10 times smaller than the original tfrecs which is insanely helpful if you ask me...

Hi @aiXander, can you please repost the link to create conditional datasets.The above link is not working.

[Johnson-yue]

Hey @Johnson-yue , were you able to train the conditional version from a pretrained unconditional one?

no, because this not support!! there are not pretrained conditional weight for finetuning