The converted model generate blur images

[TechSheng]

After the conversion, the model generates the blur images. I use the file "convert_rosinality_ckpt.py" to convert the "550000.pt"(the pretrained model on 256 provided by "https://github.com/rosinality/stylegan2-pytorch") model. In order to see the generated images by styleganv2 pretrained parameters, I modified the

MobileStyleGAN.pytorch/core/distiller.py

Line 160 in d67b3d4

img = self.student(style)["img"]

with "img = self.synthesis_net(style)["img"]" to generate some images. Here are some images below.

I don't know what caused this. Have you ever encountered this problem? @bes-dev

[bes-dev]

It works correct on my side. For exmaple 550000.pt:

[TechSheng]

Thank you for your reply! I use the latest code in your MobileStyleGAN repository. However, I can't generate sharp images. Can you share your local "convert_rosinality_ckpt.py" file with me? or the converted ckpt file? I want to find the reason.

[bes-dev]

@zuimeiyujianni I just do something like this:

• clone latest version of MobileStyleGAN.pytorch
  git clone https://github.com/bes-dev/MobileStyleGAN.pytorch.git && cd MobileStyleGAN.pytorch
• downnload 550000.pt
  gdown https://drive.google.com/uc?id=1PQutd-JboOCOZqmd95XWxWrO8gGEvRcO
• convert 550000.pt to internal format:
  python convert_rosinality_ckpt.py --ckpt 550000.pt --ckpt-mnet mnet.ckpt --ckpt-snet snet.ckpt
• create config.json:

{
    "logger": {
        "type": "NeptuneLogger",
        "params": {
            "offline_mode": true,
            "project_name": "bes-dev/stylegan2_compression",
            "experiment_name": "baseline"
        }
    },
    "trainset": {
        "emb_size": 512,
        "n_batches": 10000
    },
    "valset": {
        "emb_size": 512,
        "n_batches": 200
    },
    "teacher": {
        "mapping_network": {
            "name": "mnet.ckpt"
        },
        "synthesis_network": {
            "name": "snet.ckpt"
        }
    },
    "distillation_loss": {
        "perceptual_size": 256,
        "loss_weights": {
            "l1": 1.0,
            "l2": 1.0,
            "loss_p": 1.0,
            "loss_g": 0.5
        }
    },
    "trainer": {
        "monitor": "kid_val",
        "monitor_mode": "min",
        "style_mean": 4096,
        "style_mean_weight": 0.5,
        "num_workers": 0,
        "lr_student": 5e-4,
        "lr_gan": 5e-4,
        "batch_size": 2,
        "max_epochs": 100,
        "mode": "g,d",
        "reg_d_interval": 16,
        "truncated": false
    }
}

• run compare.py demo:

python compare.py --cfg config.json --ckpt ""

So, you should see the correct images generated by StyleGAN2 on the left (and noise on the right, because your MobileStyleGAN model is not trained yet)

[bes-dev]

@zuimeiyujianni so, we improve user experience a little bit. Just do like:

$ git clone https://github.com/bes-dev/MobileStyleGAN.pytorch.git && cd MobileStyleGAN.pytorch
$ gdown https://drive.google.com/uc?id=1PQutd-JboOCOZqmd95XWxWrO8gGEvRcO
$ python convert_rosinality_ckpt.py --ckpt 550000.pt --ckpt-mnet mnet.ckpt --ckpt-snet snet.ckpt --cfg-path config.json
$ python demo.py --cfg config.json --ckpt "" --generator teacher

It works well for me.

[TechSheng]

Thank you for your reply!!!. I have overcome this problem.