Ask for pre-trained models

[sangwoomo]

Hi, thank you for sharing your code! Do you have any plan to share your pre-trained models?

[rosinality]

Yes, I will after training is finished. But it requires a lot of time... (about 2 weeks)

[sangwoomo]

Thanks!

[linghai06]

@rosinality By the way, which dataset did you use for training. You training results seem pretty good.

[rosinality]

@linghai06 It's trained on FFHQ.

[JNUChenYiHong]

@linghai06 It's trained on FFHQ.

Thank you for your code. Looking forward to your pretraining model. The image of your training is really amazing.

[apchenstu]

Hi there, thank for this amazing reproduction! It would be really helpful if you could share your pre-train model or a more detailed train setting(e.g., number of GPUs, resolution, training time, batch size, etc.), thank you in advance!

[rosinality]

@apchenstu I have used 4 P40 with batch 32. (8 batch per GPU) If fully train 256px model with 800k iterations it will require about 15 days. I have trained about 600k iters, but FID was not satisfactory. (about 4.5)

[rosinality]

@apchenstu You can use this: https://drive.google.com/open?id=1PQutd-JboOCOZqmd95XWxWrO8gGEvRcO

[yejinyou]

Hi When I am trying to load the .pt model with torch.load in python 3.6 I am getting the following error _pickle.UnpicklingError: invalid load key, '<'. I was wondering if you could help me out? Thank you.

[caandewiel]

@yejinyou Which PyTorch version are you using? I am on 1.4.0 and it loads fine for me. Maybe try downloading the file again?

@rosinality Thanks for sharing this checkpoint. Is it correct that this checkpoint does not contain the noise weights for the synthesis network? I got it working by manually removing the noise layers, but it does not seem right to me. :)

[rosinality]

@caandewiel Yes, this checkpoints doesn't have noises as it is trained before support of fixed noises. But you can just load it with load_state_dict(strict=False).

[terrybroad]

Hi @rosinality thanks for the excellent work! Would you also consider sharing the converted model weights? I was able to convert them using your script but I do not have a 16GB GPU and I am unsure as to whether something went wrong in the conversion process, as I was getting memory errors and about 50% of the samples have some kind of artifact in them.

[caandewiel]

@terrybroad I converted them on 11GB, worked fine :) Which CUDA and PyTorch version are you using?

[terrybroad]

The machine I used for conversion had Tensorflow 1.14.0, PyTorch 1.4.0 and Cuda 10.0 using a 12GB GPU.

Maybe this is normal but a number of samples have artifacts like the ones below, do you get the same artifacts in your samples?

[caandewiel]

Hmmm, that looks interesting. Are you using truncation? I just converted the model again. Not all samples are super realistic, but none of them contain these artifacts. Some of the generated samples:

[terrybroad]

Truncation during inference in PyTorch or during conversion?

Would you mind sharing your converted weights (if possible) so I can see if I am getting the same problem with them on my end?

[rosinality]

@terrybroad You can may use this: https://drive.google.com/open?id=173WmV5EhFfMQTeDpYkLAJi3qhPtZOQs5 But I think converted weights itself will not be problematic. Samples without truncation trick could have some artifacts. You need to use truncation tricks during inference if you want visually pleasing samples.

[terrybroad]

Ah OK sorry, I missed that. I have copied your code for handling truncation in fid.py into generate.py and the samples I am getting are much better now. @rosinality would you like me to do a PR so this is handled in generate.py?

[rosinality]

@terrybroad Yes! Thank you.

[terrybroad]

Will do! Also @rosinality I am building on your code in my current research (both this repo and the original StyleGAN PyTorch implementation). How would you like to be cited if this ends up in a publication?

[rosinality]

@terrybroad I don't know it is worth to cite this repository as this is just a reimplementation or port from official repositories. But if you add footnote/cite url (https://github.com/rosinality/stylegan2-pytorch), than it would be a great honor for me. :)

[terrybroad]

@rosinality Will do! In fact, I already have cited the original StyleGAN PyTorch repo in a footnote in this pre-print: https://arxiv.org/pdf/2002.06890.pdf (page 3). But I realised I should ask for future work / camera-ready publications. If you want I can also include your real name as well in the future? :)

[rosinality]

@terrybroad If you want you can include real names, but for me url for repos is enough. Thank you.

[Daic115]

The machine I used for conversion had Tensorflow 1.14.0, PyTorch 1.4.0 and Cuda 10.0 using a 12GB GPU.

Maybe this is normal but a number of samples have artifacts like the ones below, do you get the same artifacts in your samples?

Setting t--truncation to 0.5, this is official setting in StyleGan2

[pkuanjie]

Thank you for your sharing! Could you please share the pre-trained model on the MetFace dataset?

[justimyhxu]

Can you share the training script with args? I really want to know the exact bs, lr and other hyperparameters.

[rosinality]

@justimyhxu You only need to change batch sizes so batch size * n_gpu to be 32.