small grid-like noise artifacts in the generated images.

[YadiHe]

Hello, thanks for your nice work.I have encountered an issue during my training process - I often see small grid-like noise artifacts in the generated images. I suspect this might be caused by the attention mechanism in my generator network (netG).
Could you please help me understand this problem and provide some suggestions on how to address it? I would appreciate your guidance on this matter.
Thank you.

[victorcaquilpan]

Hi @YadiHe, I am not totally sure, but the problem could be related to the initialization of the weights. Have you seen this after the training or in the beginning. And have you seen this in multiple runs?

[YadiHe]

When I train the model with 'att' for 200 epochs, I usually start to see small grid-like noise artifacts appearing around the 20th epoch, and they get worse and worse over time. However, when I switch to using 'resnet_9blocks' instead, the training seems to be progressing normally.

[victorcaquilpan]

@YadiHe, as you said, it could be a problem with the net. Just in case, your dataset is public. I would like to explore the reasons behind this.

[YadiHe]

@YadiHe, as you said, it could be a problem with the net. Just in case, your dataset is public. I would like to explore the reasons behind this.

Sorry, my data is private. Let me elaborate my background, I am training to synthesise CT from CBCT, CBCT and CT are rigidly aligned. Now I'm trying to build UNetResNetAttention, combining U-Net's jump connectivity, ResNet's deep feature extraction and multi-scale attention mechanism.