IRSE50 model for pretraining pSp on non-face datasets

[kaselby]

Hello there!

I have been very interested in trying to reimplement aspects of this project, but have had difficulty in pretraining my own pSp models. The pSp model requires the pretrained IR_SE50 model as part of its ID-loss and as an initialization for the encoder backbone. The weights given in the pSp repository were pretrained on human faces only, I believe. Did you use these same pretrained weights for training your pSp models on datasets like AnimalFaces and VGGFlowers? Or did you train a new network for these? If so, are those weights available anywhere?