Pytorch-cyclegan-and-pix2pix: continue_train

[biexiangduo]

[biexiangduo]

when i break train (i.e epoch=3),and I add "--continue_train" in command.But it failed with "no such file named lateset ..."So I just add "--epoch_count 3" 。It seems work,start from epoch 3 but the curve has break relationship of previous curve that epoch1 and 2 show .continue seem isolated and no relation with previous work.

[ghost]

You should set --continue_train --epoch 3 --epoch_count 4 (3 is the current final epoch).
--epoch option means which weights is loaded for initialization of networks.
(Default value of this option is latest.)
--epoch_count option means the start number of epoch count.

In your case, when only --continue_train was set, latest weights were supposed to be loaded because you didn't specify --epoch option.
Then "no such file" error occurred because latest weights were not saved yet.

And when only --epoch_count 3 was set, the program started training from scratch with start epoch number 3, because you didn't set --continue_train option.
That's why your curve after epoch 3 had no relationship with epoch 1 and 2.

[biexiangduo]

Thanks for update.
I wish to express my appreciation for your help.

[biexiangduo]

my last train in epoch 3 ,but not finished. Then I continue"python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan --gpu_ids -1 --epoch_count 3"

I try"python train.py --dataroot ./datasets/maps --name maps_cyclegan --model cycle_gan --gpu_ids -1 --continue_train --epoch_count 3" .But show"FileNotFoundError: [Errno 2] No such file or directory: './checkpoints/maps_cyclegan/latest_net_G_A.pth'"

should I restart from epoch 1?

[biexiangduo]

I break down train by "ctrL+c".Would this operation save training?

thanks

[biexiangduo]

I think i forget add "--save_latest_freq" and "--save_epoch_freq" so that I can continue to the latest training

[junyanz]

I am glad that you figured it out.