microsoft / Relation-Aware-Global-Attention-Networks

We design an effective Relation-Aware Global Attention (RGA) module for CNNs to globally infer the attention.
MIT License
337 stars 65 forks source link

CrossEntropyLoss & TripletLoss issue #24

Open JIaqiYang78 opened 3 years ago

JIaqiYang78 commented 3 years ago

When I trained this model, I have set up Lr=0.0008, weightdecay=0.0005 and I have checked the dim of feat and cls_feat, all of them are correct. But the CrossEntropyLoss fluctuates around 6 and it didn't reduce with epochs increasing. And the TripletLoss is 0.00000 or 0.45000 and in most cases it is 0.0000. And the prec is very low which fluctuates between 0 and 1. Does anyone know the reason for this porblem? May be we can discuss about it. Thanks.