8000 A puzzling question about the implementation of ppi · Issue #7 · gordicaleksa/pytorch-GAT · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content
A puzzling question about the implementation of ppi #7
Open
@TaengguTae

Description

@TaengguTae

Thank you @gordicaleksa for the tutorial and implementation of GAT, which made me understand it smoothly and walk through the codes sucessfully !

Here I still have a question when I run the codes of PPI dataset. In the code you have mentioned that skip connection is important and keep it otherwise the micro-F1 is almost 0, and I am wondering that the reason of the appearance of 0 is that, something leads some values(such as loss, feature etc.) to be nan during training?

More details:
The framework I used to train is paddlepaddle instead of pytorch. When I ran my code, I found that the value of loss would be nan finally, making the micro-F1 be 0. This strange phenomenon occurs whether I use skip connection or not. Furthemore, when I set skip connection to False, the value of micro-F1 first rises normally to about 0.7, then suddenly becomes 0, however, when I set skip connection to True, the value micro-F1 also rises normally first, to around 0.9, and suddenly becomes 0.
When I checked it, I found the appearance of NaN during training. This problem has always puzzled me, so I wonder that is there the same issue during your implemrntation?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions

      0