Why does the loss decrease during training, but the PSNR value gets lower?

[Joeytime]

(1) I replaced different training sets, and this problem always occurred during training. The loss has been decreasing, but the highest PSNR value is often the beginning epoch.

(2) Why is the input for training here the '../dataset/REDS/train/HR' folder?
|--train
|--HR
|--video 1
|--frame 1
|--frame 2
：
|--video 2
:
|--video n

def set_template(args):
if args.template == 'Self_Blind_VSR_Gaussian':
args.task = "FlowVideoSR"
args.model = "PWC_Recons"
args.save = "Self_Blind_VSR_Gaussian"
args.data_train = 'REDS_BLURDOWN_GAUSSIAN'
args.dir_data = '../dataset/REDS/train'
args.data_test = 'REDS_HRLR'
args.dir_data_test = '../dataset/REDS4_BlurDown_Gaussian'
args.HR_in = False
args.scale = 4
args.patch_size = 160
args.n_sequence = 5
args.n_frames_per_video = 50
args.n_feat = 128
args.extra_RBS = 3
args.recons_RBS = 20
args.ksize = 13
args.loss = '1*L1'
args.lr = 1e-4
args.lr_decay = 100
args.save_middle_models = True
args.save_images = False
args.epochs = 500
args.batch_size = 8
If I replace it with LR, there will be a patch size problem.

[qqhe-frank]

I having the same first problem where the losses are going down, but the metrics are going down just as much. I guessing the reason for this is that the reconstruction network doesn't really learn during the training process, but rather amounts to a residual shortcut that leads to worse and worse PSNR results.

[SaeedNowrouzian]

I am having trouble installing and running this code. If possible, friends who have successfully implemented the code. Write the requirements.txt file and type of operating system.