rosinality / style-based-gan-pytorch

Implementation A Style-Based Generator Architecture for Generative Adversarial Networks in PyTorch
Other
1.1k stars 233 forks source link

checkpoint question #103

Open RihardsVitols opened 4 years ago

RihardsVitols commented 4 years ago

dear all,

can some one tel me how i can load checkpoint in to train.py?

i was trying this: python train.py --mixing lmdb --ckpt checkpoint/010000.model

resoult: File "train.py", line 321, in generator.module.load_state_dict(ckpt['generator']) KeyError: 'generator'

if i try with out .model file i get: PermissionError: [Errno 13] Permission denied: 'checkpoint'

thank you

rosinality commented 4 years ago

Only train_step-X.model checkpoints have generator and discriminator weights. Other checkpoints only have weight of the ema of the generator.

RihardsVitols commented 4 years ago

Thank you.

and thank you for sharing this code.

quyet0nguyen commented 3 years ago

@rosinality i don't know why the purpose why you save the 010000.model checkpoint and train_step-X.model. Because when I trained model after a break, the model started training from the begining (from 8x8 resolutions) instead of saved state dict.

Thank for your code very much.

rosinality commented 3 years ago

@quyet0nguyen You can use --init_size argument to beginning for more larger resolutions.

quyet0nguyen commented 3 years ago

@rosinality thanks for your answer.