There are some questions about the code and can not reproduce your work

[DON738110198]

首先感谢您的优秀的工作，但是我在运行的时候仍然发现了一些令人费解的问题：
in here：

OpenPSG/kings_sgg/models/relation_heads/relation_transformer_head_v4.py

Line 76 in 3ea4d68

img_size=None, patch_size=patch_size, in_chans=object_feature_size, embed_dim=object_feature_size)

我认为这个img_size不应该被设置为None，因为这会导致无法正常初始化，在后续初始化时报错。如下图所示：

如果我传入一个固定的img_size比如224，虽然可以正常初始化和训练了（但是训练时传入的图像尺寸并非固定为224），但是训练的效果会出现问题，我和我的老师怀疑是因为图片尺寸没有固定而导致训练时的ground truth无法正常对齐而导致的。

还有就是我训练的时候会出现grad_norm极大的情况，达到几千万上亿！正常来说应该需要对其进行裁剪，限制其上限，不知道为什么这份代码中没有这种操作，能否让我看看您的训练日志呢？

First of all thank you for your excellent work, but I still found some puzzling problems when running:
in here: https://github.com/franciszzj/OpenPSG/blob/3ea4d680becf55d64237cfe66d7d30640ef87d6f/kings_sgg/models/relation_heads/rel ation_transformer_head_v4.py#L76

I don't think this “img_size” should be set to None because it will not initialize properly and will cause errors in subsequent initializations.

If I pass in a fixed img_size such as 224, although it can be initialized and trained normally (but the image size is not fixed to 224 during training), the training effect will be problematic. My teacher and I suspected that it was because the picture size was not fixed, so the ground truth in the training could not be properly aligned.

Also, when I train, there will be a grad_norm extreme situation, reaching tens of millions of millions! Normally, it should be cropped to limit its upper limit. I don't know why there is no such operation in this code. Could you please show me your training log?

[hanzefang]

@DON738110198 Do you solve the issue ?We also met the question.

[rafa-cxg]

is this issue solved?