weijiawu / syntoreal_std Goto Github PK

View Code? Open in Web Editor NEW

34.0 34.0 7.0 304 KB

HHH

Python 100.00%

syntoreal_std's People

Contributors

Stargazers

Watchers

Forkers

zhengjiawen clscy cv-ip doem97 thorpham shearylee omeremcgehee

syntoreal_std's Issues

Mask all the groudtruth (for target domain, the pseudo label)

SyntoReal_STD/dataset/ICDAR15.py

Line 330 in 02ee312

 ignored_polys.append(np.around(scale * vertice.reshape((4, 2))).astype(np.int32)) 

Sorry but I think this is really not a code that can be followed up with. In this line, the data loader for ICDAR15 masked all the predictions which troubled me for a long while :(

I'd recommend followers wait for further updates or bug fixes from the author, otherwise, you may run into random logic bugs. If you insist, be reminded to fix the above prob.

Besides, I will also release a related but not exactly re-implement of syn2real later.

Trained weights?

Thanks for your amazing work, When the trained weights will be available to test?

Train Model with new language

It's a amazing model and I want to train it with new language. How can I do that?

where is the "./evaluate/script.py"?

Dear author, could you please release the complete code? Thanks!!!!!!!!!!!!!!!!!!

Uuu

Can you upload weight on google driver ?

Thank you for your work . The paper is great and i want to test model but i can't download baidu server. Please upload weight on google driver . Thank a lot

What's the source and target domain for the pretrained model?

Hi @weijiawu Mr Weijia Wu, thanks for your work and code sharing!

May I know what is the source/target domain for this pre-trained model? (e.g. Synthtext to ICDAR 15 or 13?) I can not re-produce your results in the paper.

Thanks for any useful information!

Where is 'network/loss_target.py' file?

Thank you for your great work!

In the training file, 'network/loss_target.py' is needed to to import Loss_target, but i cannot locate it.

Would you update that part?

Thanks.

如何用ATA模块去预训练模型？

你好，请教您一个问题：如何利用ATA方法去学习域无关的特征呢？就是论文中的第三个公式是怎么实现的呢？

Any plans to release the training code?

Hi authors,

Thank you for your great work. Will you be releasing the training code? It's been "TBD" for quite a while.

Best,

where is 'net work/loss_target.py' file .请问您有这一个文件可以给我吗，这对我非常重要，非常感谢！！！Do you have this one document for me, which is very important to me, thank you very much

Loss function parameters mismatch

Hi, the loss function in train() and Loss() have a different number of parameters. (5 params vs 6 params)

Besides, the Loss_target() Class is also missing (in train.py). Are you using a different loss lib other than the provided one? Thanks!

为什么不在每个训练epoch中更新pseduo label ?

SyntoReal_STD/trainSyndataToICDAR15.py

Lines 138 to 143 in 02ee312

 generate_pseduo(model, args.target_image, args.target_pseudo_positive, args.target_pseudo_negative, device) 

 for epoch in range(args.epoch_iter): 

 train( epoch, model, optimizer, train_loader_source,train_loader_target) 

 # 进行target domain的eval，看看指标。

可以改成以下这样吗？

for epoch in range(args.epoch_iter):
    generate_pseduo(model, args.target_image, args.target_pseudo_positive, args.target_pseudo_negative, device)
    train(epoch, model, optimizer, train_loader_source, train_loader_target)

	generate_pseduo(model, args.target_image, args.target_pseudo_positive, args.target_pseudo_negative, device)

	for epoch in range(args.epoch_iter):

	train( epoch, model, optimizer, train_loader_source,train_loader_target)
	# 进行target domain的eval，看看指标。