How to use the data produced by the build_dataset.py file during training

[clovaai]

How to use the data produced by the build_dataset.py file during training

[zj916716524]

How to define unseen_unis and seen_unis, unseen_fonts and seen_fonts in the 'valid' field of the tarin.json file

[zj916716524]

I have rewritten it according to the format given by train.json, combined with my own dataset.
unseen_unis : some characters that are not visible during training
seen_unis : I define all the characters that are seen in the training set
unseen_fonts: invisible fonts for training
seen_fonts: fonts used during training
I don't know if this is the correct way to write it. And about the avail, I defined the fonts seen and unseen in training and the characters contained in the fonts.

[8uos]

Hi, sorry for the late reply.
Firstly, I recommend to use this repository (clovaai/fewshot-font-generation) because that repo builds the dataset from TTF files -- that means, you do not need to build the LMDB datasets.

However, you are doing right to build the dataset for this repository.

[zj916716524]

Thank you very much for your reply, I followed the data set I made myself and trained the network, the network returned a BUG
{
File "/home/zeng/for_translation_stroke/lffont-master/datasets/p1dataset.py", line 129, in
content_imgs = torch.cat([self.env_get(self.env, self.content_font, uni, self.transform)for uni in trg_unis]).unsqueeze_(1)
File "train.py", line 141, in
env_get = lambda env, x, y, transform: transform(read_data_from_lmdb(env, f'{x}_{y}')['img'])
TypeError: 'NoneType' object is not subscriptable

}
Source code：
{
content_imgs = torch.cat([self.env_get(self.env, self.content_font, uni, self.transform)for uni in trg_unis]).unsqueeze_(1)
}
I don't know what caused this bug. Is it related to the insufficient memory of the graphics card?
Thanks a lot for the link, so should I build the dataset from this file?

[zj916716524]

I ran the code of your other article MX-FONT. Can the method of making a dataset in that article be applied to LF-FONT? The network ran successfully when I made the MX-FONT dataset, and got exciting results.

[8uos]

Thank you very much for your reply, I followed the data set I made myself and trained the network, the network returned a BUG { File "/home/zeng/for_translation_stroke/lffont-master/datasets/p1dataset.py", line 129, in content_imgs = torch.cat([self.env_get(self.env, self.content_font, uni, self.transform)for uni in trg_unis]).unsqueeze_(1) File "train.py", line 141, in env_get = lambda env, x, y, transform: transform(read_data_from_lmdb(env, f'{x}_{y}')['img']) TypeError: 'NoneType' object is not subscriptable

} Source code： { content_imgs = torch.cat([self.env_get(self.env, self.content_font, uni, self.transform)for uni in trg_unis]).unsqueeze_(1) } I don't know what caused this bug. Is it related to the insufficient memory of the graphics card? Thanks a lot for the link, so should I build the dataset from this file?

File "train.py", line 141, in
env_get = lambda env, x, y, transform: transform(read_data_from_lmdb(env, f'{x}_{y}')['img'])
TypeError: 'NoneType' object is not subscriptable
}
This error usually caused by that the given character does not exist in the given font.
Please check your meta file carefully.
If you ran build_dataset.py, you can find the json file which contains the available font-char dictionary at --json_path.

[8uos]

I ran the code of your other article MX-FONT. Can the method of making a dataset in that article be applied to LF-FONT? The network ran successfully when I made the MX-FONT dataset, and got exciting results.

The repository clovaai/fewshot-font-generation uses very similar dataset code with MX-Font repository and it also supports LF-Font.
It will be much easier to use because it does not need to build LMDB dataset and also the meta file for that repository is much simpler.
Please check the repository and the docs/Dataset.md.

[zj916716524]

Thank you very much for your quick reply. I will modify the project according to your suggestions next. If I have a problem, I'm here to ask you.

[SanghyukChun]

Closing the issue, assuming the answer resolves the problem.
Please re-open the issue as necessary.