Can I change a custom resnet 18 architecture subtly and still use it in pre-trained mode?

[Mona_Jalal]

or for example should I use this answer?

Missing keys & unexpected keys in state_dict when loading self trained model

I also am facing the same issue. I don’t use nn.DataParallel. I am using a slightly modified version of [this repo] in a Kaggle notebook https://github.com/aitorzip/PyTorch-CycleGAN. Here’s how I save: torch.save({ 'epoch': epoch, 'model_state_dict': netG_A2B.state_dict(), 'optimizer_state_dict': optimizer_G.state_dict(), 'loss_histories': loss_histories, }, f'netG_A2B_{epoch:03d}.pth') Here’s how I load: checkp…

Can you at a high-level (not code-level) please provide some hint why this answer should work?

I am also loading a torch model like below:
checkpoint = torch.load(args.resume, encoding='latin1')