Unable to replicate results from using torch.hub

vision
1

I am trying to generate celeb faces using pretrained PGAN from torch.hub, however I cant seem to get decent images at all. I am unsure what I am doing wrong. Here is what I did

import torch
hub_model = torch.hub.load(
    'facebookresearch/pytorch_gan_zoo:master',
    'PGAN', #
    config = None,
    useGPU = True,
    pretrained=True)

inputRandom, randomLabels = hub_model.buildNoiseData(16)
im = hub_model.test(inputRandom,
           getAvG=True,
           toCPU=True)

The images I get dont look as great as the ones shown on https://github.com/facebookresearch/pytorch_GAN_zoo.
This is the image i generated

image
image1570×1096 191 KB

Are the images there just cherry picked or am I doing something wrong?

2

Did you try to generate a few more faces?
The last one looks quite good and some faces in the repo docs look also quite weird, so you might just have gotten some randomly bad results.

3

These are randomly generated results

image
image1662×1184 418 KB

It is kind of 50/50, so im not sure this is as good as the paper’s.