search

LynnHo / AttGAN-Tensorflow

AttGAN: Facial Attribute Editing by Only Changing What You Want (IEEE TIP 2019)
MIT License
603 stars 133 forks source link

About design of model_348 #53

Closed NoctisZ closed 2 years ago

NoctisZ commented 2 years ago

Thank you very much for sharing the code and result! I noticed that for the model_348, the dimension of Conv layers is reduced (64 -> 48) and that of the FC layer is also set to half (1024 -> 512) compared with model_128. I'm pretty new to GAN and wonder if there's any specific reason(s) behind this hyperparameter design? Is it mainly because we want to speed up training by reducing the number of weights? Any information will be greatly appreciated.

LynnHo commented 2 years ago

@NoctisZ Just because I got OOM on my machine keeping conv-64 and fc-1024😭

NoctisZ commented 2 years ago

@LynnHo I see, that's reasonable haha. Thanks for the info!