Using basicvsr_pp.py with another training setting

[NasrinR791]

Hi @ckkelvinchan

I can train basicvsr++ with default settings using mmediting codes without any problem and it fits in the gpu memory.
However, when I use https://github.com/xinntao/EDVR/tree/old_version to train basicvsr_pp.py as the model (replacing BasicVSR++ with the EDVR model), the training stops with cuda memory error even with batch size of 1.

In https://github.com/xinntao/EDVR/tree/old_version in codes/models/networks.py I specify:

from mmedit.models.backbones.sr_backbones.basicvsr_pp import BasicVSRPlusPlus
if which_model == 'BasicVSR_pp':
netG = BasicVSRPlusPlus(mid_channels=64,
num_blocks= 7,
max_residue_magnitude= 10,
is_low_res_input= False,
spynet_pretrained = opt_net["spynet_pretrained"],
cpu_cache_length= 100)

Do you have any idea what is the difference between training with https://github.com/xinntao/EDVR/tree/old_version and mmediting software?

Thank you for your help in advance

[zengyh1900]

Please check this issue. @Z-Fran

[Z-Fran]

Sorry for the late reply. Can your supply logs of training?