No optimizer for metric_fc?

[fyang93]

As far as I know, the centers (self.W) in losses should be updated during training. To do that, it requires using optimizers on both the model and the metric_fc (e.g. ArcFace). However, I did not find any optimizer set on metric_fc in your code so technically the centers won't be updated anyway... Is it supposed to be like that?

[melgor]

Very good catch :) Then only backbone is learned and in classification layer just random weight stay for all training. Author should change this implementation, guided by your explenation.

[Atharva-Phatak]

what you can do is inside your model class define let's say you are using ArcFace. define it as follows

class Model:

   def __init__(self , args):
         #Define your model here
        self.fc = ArcFace(num_features , n_classes)

 def forward(self , inp , targets):
        feature = self.model(inp)
        arc_op = self.fc(feature , targets)
        return arc_op

when you define optimizer opt = SGD(model.parameters() , 1e-3)

I think this should work, since the ArcFace Repo you mentioned gets only features(in that repo 512 dimensional embedding) from the Resnet Model and the passes it to ArcFace module. Hence I think defining model as above should in theory work.