Some Question About The SVD Decomposition

[yuanlonghui]

Congratulations! It's not only a marvelous idea for few-shot learning, but also a feasible idea for parameter efficient adaptation. So, I tried to apply SVD decomposition for a source pretrain model, then adapt it using self-training with the pseudo label under the source-free scenario. But I met the problem that the loss always produce tiny gradients, which is almost invaild for adaptation. So I turned to you for help. Have you ever encountered this problem？Waiting for your early replay.

[syp2ysy]

Thank you for your interest, I didn't fully understand your question, can you provide more details?

[yuanlonghui]

Can we communicate via email or other private communication channels?

[XXING1999]

Why do I report an error when I reproduce it. Like ，

[XXING1999]

and this . it reminds there is an error ，nn no has conv2d attribute

[syp2ysy]

@yuanlonghui You can provide some details in the email, hoping to help you clear your doubts. My email is yanpeng_sun@njust.edu.cn

[syp2ysy]

@XXING1999 sorry, you can change nn.Conv2d to nn.Conv2D.

[XXING1999]

上面的那个resnet没有name_modeule属性已解决，是因为我采用了你的测试代码，使用的paddle框架，里面resnet继承的是nn.layer属性，不是nn.module. 请问一下这个代码有pytorch框架实现的吗

[syp2ysy]

@XXING1999 你邮件联系我一下，我邮件发给你

[XXING1999]

非常感谢，邮件已发，望您查收。

[syp2ysy]

@XXING1999 我没有收到邮件啊？请确认邮箱地址是否正确。

[XXING1999]

抱歉，我又重新发送了一下，您再查看一下

[zhangyuxuan321]

@XXING1999 你邮件联系我一下，我邮件发给你 可以麻烦也发我一份吗，我的邮箱是815594811@qq.com ,感谢！