About resume training

[zyfan1]

if args.restore_ckpt is not None:
    strStep = os.path.split(args.restore_ckpt)[-1].split('_')[0]
    total_steps = int(strStep) if strStep.isdigit() else 0
else:
    total_steps = 0

Hi,
I found this code in train.py. The class OneCycleLR need a paramter 'last_epoch' to resume the lr, so it needs save the optimizer's state_dict, because we need a key called 'initial_lr'. However, I don't find the code to save optimizer's dict. Why?
Do you have to rename the model file when you load it? This would allow this code to be skipped.

[xiaofeng94]

Hey,
We didn't save optimizer's state_dict. The total_steps is attained from the file name (usually it ends with the number of iterations). So when resume training, you may expect some differences from your original training process.

[zyfan1]

Thank you for replying，
if 'initial_lr' not in group:
raise KeyError("param 'initial_lr' is not specified "
"in param_groups[{}] when resuming an optimizer".format(i))
Look at this code in OneCycle line 32. My question is there are no codes to add the parameter 'initial_lr' to optimizer in train.py, so if the 'last_epoch' in OneCycle is set, wouldn't an error be thrown at runtime?

[xiaofeng94]

Hey, why not take a try. I think initial_lr may be set somewhere in RAFT's codebase.

[zyfan1]

How do you compare the results of your experiments with the 2-view ones when you are warm-start?

[xiaofeng94]

Hey, we only report the 2-view results. You may check the caption of Table 2 in the paper or run the experiments on your own.