Unable to test your own training results

[LVCHENYONG]

Hello author, I trained my own style migration dataset, but output shadow filling when testing. I would like to ask what is going on, where the setting is wrong?

[usert5432]

Hello @LVCHENYONG,

I'm sorry to hear about the shadow filling issue in your style migration dataset.

GAN training has many pitfalls and without knowing specific details, diagnosing the cause will be very challenging. I was wondering if it would be possible for you to provide the training script you have used and a few samples of failed translations? This could help us to diagnose the problem.

[LVCHENYONG]

Good morning, I was training the style transfer from A to B, but typing A during the test got a shadow-filled face, not B.

This is my pre-training script:
python scripts/train/selfie2anime/bert_selfie2anime-256.py

This is my training script:
python scripts/train/selfie2anime/cyclegan_selfie2anime-256.py --attr selfie2anime

This is what I tested:
python scripts/translate_images.py outdir/selfie2anime/model_d(cyclegan)_m(autoencoder)_d(None)_g(vit-unet)_bert-vit-unet-12-256 -n 100000

[usert5432]

Hi @LVCHENYONG,

Thank you for elaborating on the issue! Looking at the commands:

python scripts/translate_images.py outdir/selfie2anime/model_d(cyclegan)_m(autoencoder)_d(None)_g(vit-unet)_bert-vit-unet-12-256 -n 100000

I think you may be testing a BERT-like pre-trained model, instead of a style migration model. Could you please check if there is another model saved in the outdir/selfie2anime/ directory?

[LVCHENYONG]

Thank you for your answer. I didn't pay attention. I used the self-supervised weights