About figure of features

[zorrocai]

I wonder that how to change the code to plot the figure of features with only softmax loss.
And what do you think about the different shape between the Fig.3 in original paper and the figure produced though your codes.

[jxgu1016]

1. set loss_weight to 0
2. I guess that difference may be caused by the initialization of the centers. They may initialize them around a circle.

[zorrocai]

I tried the circular initialization, and it doesn't match your guess.

[jxgu1016]

Sorry to hear that. There may be other tricks we don't know..

[zchky]

@zorrocai

set the bias to False, you can get the circle

[jxgu1016]

Interesting, did you find the solution yourself on this code or refer to other issues on other repo? @zchky

[zchky]

@jxgu1016

you can find something on the paper, (normface: l2 hypersphere embedding for face verification)

[zchky]

more specifically

Drop the bias of last fully connected layer.

[jxgu1016]

See Fig. 3 in the paper of NormFace.