search

junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch
Other
22.4k stars 6.22k forks source link

Generate person image from the segmentaion one (COCO dataset) #1288

Open QLinhHub opened 3 years ago

QLinhHub commented 3 years ago

Hello, i have a problem when doing image translation from SEGMENTATION to COLOR IMAGE using Pix2Pix. I have > 4000 images for training, my batch size is 16 and my learning rate is 0.0001. After 200 epochs, the model generate well on training set, as bellow: image But on test set, i have something like: image Is it the overfitting problem? If yes, what should i do in this case?

  1. I think about add some regularizer layers but what value of lambda should i set?
  2. Should i reduce my batch size to 1 or just use Instance Norm? By the way, i'm working on PERSON DEIDENTIFICATION problem, is anyone working on this topic? Every answer will be appreciate! Thanks in advance!
QLinhHub commented 3 years ago

Hello, can someone help me please !!!!

synthetica3d commented 2 years ago

Try using --norm instance because batch norm gave me endless headaches, and I found an issue with the batchnorm's behaviour after eval().