seems pretrain model is not used

[Yang Jiao]

Hi, nice work, appreciate it! But two questions confuse me.

1. in train_vtoonify_d.py, for pre-training, why save the weights of g_ema in line 172 / 387? Looks g_ema.eval() keep the weights unchanged, and g is the generator should to be trained.

2. after pre-training, in full train process, looks the pre-trained model is not loaded (from vtoonifu_d_cartoon/pretrain.pth).

Thanks!

[Shuai Yang]

This part updates g_ema with g, so its weight changes.

VToonify/train_vtoonify_d.py

Lines 340 to 342 in db57c27

 accumulate(g_ema.encoder, g_module.encoder, accum) 
 accumulate(g_ema.fusion_out, g_module.fusion_out, accum) 
 accumulate(g_ema.fusion_skip, g_module.fusion_skip, accum)         

This part loads pre-trained model

VToonify/train_vtoonify_d.py

Lines 68 to 69 in db57c27

 if self.opt.encoder_path is None: 
     self.opt.encoder_path = os.path.join('./checkpoint/', self.opt.name, 'pretrain.pt') 

VToonify/train_vtoonify_d.py

Lines 430 to 431 in db57c27

 if not args.pretrain: 
     generator.encoder.load_state_dict(torch.load(args.encoder_path, map_location=lambda storage, loc: storage)["g_ema"]) 

[Yang Jiao]

thanks a lot :D appreciate it ~

one more question, could you please provide the full pipeline for training all nets from scratch? just a brief flowchart works fine for me, just a little bit confusing since training ALL-models looks tough. :(

[Shuai Yang]

we only has two steps
pretrain the encoder (train the green part) and train the full network (train the green and blue parts).

[Yang Jiao]

Thanks! By the way, what does ema in g_ema mean? (non-abbrevation version is?)

[Shuai Yang]

exponential running average
rosinality/stylegan2-pytorch#34

[Yang Jiao]

thanks a lot!