用celeba_256数据集从头训练为什么无法收敛？

researchmm / AOT-GAN-for-Inpainting

[TVCG'2023] AOT-GAN for High-Resolution Image Inpainting (codebase for image inpainting)

https://arxiv.org/abs/2104.01431

Apache License 2.0

425 stars 68 forks source link

用celeba_256数据集从头训练为什么无法收敛？ #12

Closed Cynicarlos closed 8 months ago

Cynicarlos commented 9 months ago

感谢作者的贡献！请问我用celeba_256从scratch开始训练，但每张图片对应一张固定的mask，且输入图含噪声，算loss时用的是clean的GT，batch_size调成16，iteration为1e6，为什么一直没有收敛呢？最后的log如下： mt:7.5s, dt:0.0s, L1:0.390, Style:0.700, Perceptual:0.389, advg:0.006, advd:0.000, : 100%|█████████▉| 999666/1000000 [139:05:29<04:09, 1.34it/s]