search

Zhongdao / gcn_clustering

Code for CVPR'19 paper Linkage-based Face Clustering via GCN
MIT License
360 stars 86 forks source link

predicting edges are all zeros during training #7

Open jxyecn opened 5 years ago

jxyecn commented 5 years ago

Hello, when I was trying to train the model with your example, I found predicting edges came to be zeros. (No edge is predicted to be true) Have you ever met this situation?

This status usually occurs after 100 batchs' training, with following args.

k_at_hop = [200, 10]
active_connection = 10
batch_size = 16
momentum = 0.9
weight_decay = 1e-5
lr = 1e-5
Zhongdao commented 5 years ago

I suggest you check the positive/negative node number in an IPS. If the number of negative nodes is much larger than the number of positive nodes, the network is prone to bias.

jxyecn commented 5 years ago

I suggest you check the positive/negative node number in an IPS. If the number of negative nodes is much larger than the number of positive nodes, the network is prone to bias.

After I retested on CASIA-WEBFACE dataset, for the first 200 batchs, negative-positive proportion ranged from 1.5769 to 15.0804, while the mean proportion is 6.1605. May I ask if this proportion is "the number of negative nodes is much larger than the number of positive nodes"? So, how to build such a feasible train set?