StyleMapGAN - Official PyTorch Implementation

Hello authors,

I have an error when using the pretrained checkpoints when torch.load(args.ckpt) is run in generate.py (or any piece of code with the function). I tried re-downloading the model as proposed in this issue. The only change I made to accommodate my CUDA 11 version was to instal pytorch with conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch. Thanks for you attention, it is much appreciated.

Hello, thanks for sharing your code. I'm very interested in your work and I got a question when I run the metrics for reconstruction. The output of metrics.reconstruction are 0.015(mse) and 0.214(lpips). The output seems much better than the results reported in Table 3 in the paper (mse=0.024, lpips=0.242). I would appreciate it if you could give me an explanation.

I downloaded the pre-trained networks using the updated download.sh files, and pytorch is unable to load them, giving the following error: _pickle.UnpicklingError: invalid load key, ‘<‘

I suspect that the pre-trained network on the google drive is corrupted. Could you check whether it is the case?

Thanks for the excellent work!

I am failed to run the demo on my local machine. The app does not produce the wrong information and the website seems can not load the demo images. Can you help me to fix it?

I can run all the commands except the demo one such as the unaligned transplantation command.

Hello, I would first like to thank you for sharing your work.

I am having problem on loading weights from checkpoints(i.e on continuing halted training)

I am training StyleMapGAN on custom dataset(~200K images in the training dataset, 1024*1024 resoulution), and I am currently using 3 TitanRTX GPUs. I am using latent_spatial_size=16 considering input image resolution and GPU memory. On training with such configuration, batch 2 is allocated per GPU using ~21 GiB memory.

There is no problem on training from scratch. I have not tried using pretrained weights trained on FFHQ or CelebA because my data is quite different from human faces. Moreover, as I have succeeded on generating images from generate.py, I think weights were saved in proper way.

However, memory allocation problem occurs every time I load custom weights to continue training. I assumed extra memory may be required on loading weights, so I tried using smaller batch size (batch 2 per GPU->batch 1 per GPU), but same memory shortage problem occurs.

To summarize, I cannot load weights to continue training, whereas training from scratch or loading weights to generate images are working well. Thereafter, I would like to ask following questions.

1. Had any of the authors experienced with similar problems?
2. Would there be any possible solutions to my problem?

I would be grateful if you take a look into my question. Thank you!

Thank you for publishing the code. Your work is very impressive. I wonder how many iterations do you use to train the model？I noticed that the default is 1,400,000. Is this value for all your training datasets? 1400000 training iterations a bit too long, I am curious if there is a small number of training iteration.

there is how to get pretrained network

# Download raw images and create LMDB datasets using them
# Additional files are also downloaded for local editing
bash download.sh create-lmdb-dataset celeba_hq

# Download the pretrained network (256x256) 
bash download.sh download-pretrained-network-256 celeba_hq # 20M-image-trained models
bash download.sh download-pretrained-network-256 celeba_hq_5M # 5M-image-trained models used in our paper for comparison with other baselines and for ablation studies.

# Download the pretrained network (1024x1024 image / 16x16 stylemap / Light version of Generator)
bash download.sh download-pretrained-network-1024 ffhq_16x16

but with these networks, it doesn't work

File "demo.py", line 192, in <module> ckpt = torch.load(args.ckpt) File "/root/anaconda3/envs/stylemap/lib/python3.6/site-packages/torch/serialization.py", line 608, in load return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args) File "/root/anaconda3/envs/stylemap/lib/python3.6/site-packages/torch/serialization.py", line 777, in _legacy_load magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, '<'.

and i found that after run train.py, i got 000000.pt (didn't trained) and with this 000000.pt, demo.py works well,(but output image is noisy image)

so is there any way get new pretrained network?

i tried on (pytorch= 1.4.0, 1.10) and (remote server-docker , colab)

Hi，Thanks for your great work. As you say in the paper, the generator and the encoder are trained jointly not separately, how about trained together as a Unified framework without fixing weights.

how can u get MSE, LPIPS of image2stylegan, structured noise and stylegan2 in table3, i mean, structured noise didn't provide the code about how to get image reconstruction, and image2stylegan didn't provide codes(No official code found), and what about stylegan2, can u tell me please?

Hello, thank you for interesting work! used python demo.py --ckpt expr/checkpoints/celeba_hq_8x8_20M_revised.pt --dataset celeba_hq I try to evaluate quality of real photos embedding.

It works impressively good on provided by default images:

however drastically worse on photos from internet: (photos are aligned by face landmarks and cropped to 1024x1024 by https://github.com/ZPdesu/Barbershop/blob/main/align_face.py)

what can be the reason and how do I fix this?

source images:

Hello. Thanks for this code. I have trained the model on my own dataset that is different from the face dataset. The image reconstruction, random generation works well. However, If I do the style mixing their is a hard coded pkl file required that is related to the celeba_hq dataset (data/celeba_hq/local_editing/celeba_hq_test_GT_sorted_pair.pkl).

Just for the sake of running stylemixing I downloaded the dataset and ran the code, however, the results are not good.

Could you please share your thoughts on how to apply stylemixing on a custom dataset?

Hi,

I use the provided prepare_data.py to transfer the downloaded raw images to the corresponding lmdb data with 256x256 size. But I find that the lmdb data has different sizes compared to the downloaded lmdb data (LMDB_train/test/val). I also tried other lmdb sizes such as 128, 512 and 1024 but no image size can match the downloaded lmdb data.

May I figure out the reason?

Thanks.

Hi, thanks for your excellent work!

May I know how to calculate or obtain GT_labels and LMDB_test_mask for other images in CelebA_HQ as this repo just provides the processed masks for downloading via download.sh.

Thanks!