Abnormal training results

[diamond0910]

Thank you for releasing the code. I train the code using the given celeba dataset. But I get abnormal training results.

I use the default script pix2pix3D/train_scripts/celeba_seg.sh.

The fakes0000012.jpg seems good.

But the fakes000080.jpg is abnormal.

Can you give me some advice? Thank you!

[dunbar12138]

Thanks for pointing it out.

It's possibly because there was extra space at the end of some lines in the pix2pix3D/train_scripts/celeba_seg.sh. So some of the training options are not correctly read.

Could you try the newly updated training script?

[diamond0910]

Thank you for your reply. In fact, I couldn't run the original script due to formatting issues. I put them all in one row.

This is my script.

python train.py --outdir=logs --cfg=celeba --seg_weight=2 --data=./CelebAMask/train --mask_data=./CelebAMask/train_label --data_resolution=512 --render_mask=True --dis_mask=True --neural_rendering_resolution_initial=128 --gpus=4 --batch=4 --mbstd-group=1 --gamma=1 --gen_pose_cond=True --resume=./pretrained_models/ffhqrebalanced512-128.pkl --random_c_prob=0.5 --only_raw_recons=True --lambda_cross_view=1e-4 --lambda_d_semantic=0.1 --lambda_lpips=1 --wandb_log=False --snap=5 --no_eval=True

Is there any problem? Thank you.