How to improve the results

[NasrinR791]

Hi @andreas128
Thanks for your great work and sharing the code.
I tried to train SRFlow using REDS dataset. After training the RRDBN network on REDS (with PSNR around 29 dB), I used this config to train SRFlow. However, the resulted SR images are not good and the PSNR is around 17 dB.
I did not change the code by the way. Could you please give me a hint on how can I improve the training?

network_G:
which_model_G: SRFlowNet
in_nc: 3
out_nc: 3
nf: 64
nb: 25
upscale: 4
train_RRDB: false
train_RRDB_delay: 0.5

flow:
K: 10
L: 3
noInitialInj: true
coupling: CondAffineSeparatedAndCond
additionalFlowNoAffine: 2
split:
enable: true
fea_up0: true
stackRRDB:
blocks: [ 1,5, 15, 22 ]
concat: true

train:
manual_seed: 1386
lr_G: !!float 1e-5
weight_decay_G: 0
beta1: 0.9
beta2: 0.99
lr_scheme: MultiStepLR
warmup_iter: -1 # no warm up
lr_steps_rel: [ 0.5, 0.75, 0.9, 0.95 ]
lr_gamma: 0.8
weight_l1: 0.1

niter: 300000
val_freq: !!float 5e3

validation settings

val:
heats: [ 0.0, 0.5]
n_sample: 1