search

XPixelGroup / HAT

CVPR2023 - Activating More Pixels in Image Super-Resolution Transformer Arxiv - HAT: Hybrid Attention Transformer for Image Restoration
Apache License 2.0
1.2k stars 147 forks source link

Loss does not converge #110

Open sangjun04 opened 11 months ago

sangjun04 commented 11 months ago

First of all, thank you so much for a great work. really appreciate it.

I am trying to fine tune the train_HAT-L_SRx4_finetune_from_ImageNet_pretrain model with a custom data,but the loss is not converging at all.

I am using the default setting,

**# general settings name: train_HAT-L_SRx4_finetune_from_ImageNet_pretrain model_type: HATModel scale: 4 num_gpu: auto manual_seed: 0

dataset and data loader settings

datasets: train: name: Custom_data type: PairedImageDataset dataroot_gt: ../train/hr dataroot_lq: ../train/lr_deblurred_130000 meta_info_file: meta_info_custom.txt io_backend: type: disk**

and trained for over 100000 iterations, and the results are getting worse.

I am 100 percent sure there is nothing wrong with my custom dataset, since I used the same dataset to train SwinIR and the train loss converged well.

Any idea what might have gone wrong?

abdullahbas commented 4 months ago

Same here. Fine tuned model is worse than the pretrained one. I think either something wrong with the source code or they used different loss function.