Problems that arise when training gmn models with other features #58
Description
dear author：
I hope this message finds you well. I am currently working on retraining the Graph Matching Networks (GMN) model using different feature sets and have encountered a challenge that I hope you can help me with.
Specifically, I have noticed that during retraining, the model exhibits a strong bias towards the positive class. This results in the Euclidean distance between the encodings of function pairs consistently being less than 1. However, according to the loss function, the Euclidean distance for encodings of negative class function pairs should be greater than （1+margin）.
I am trying to understand the potential reasons for this bias and how to adjust the training process or feature selection to correct it. Could you provide any insights or suggestions on how to address this issue?
Thank you very much for your time and assistance.