what is the learning rate when start self-critical training?

[siyilingting]

Hi, ruotian. Thank you for your job. I am a bit confused about the learning rate (5e-5) for self-critical training.
First, I train without self-critical, set --learning_rate 5e-4 --learning_rate_decay_start 0 --scheduled_sampling_start 0 --max_epochs 30 as you write. Then after 30 epochs I train using self-critical with the setting --learning_rate 5e-5 --start_from log_fc_rl, no learning_rate_decay_start and scheduled_sampling_start.
In the train.py, I find that

self-critical.pytorch/train.py

Line 82 in 275e22c

if vars(opt).get('start_from', None) is not None and os.path.isfile(os.path.join(opt.start_from,"optimizer.pth")):

if vars(opt).get('start_from', None) is not None and os.path.isfile(os.path.join(opt.start_from,"optimizer.pth")): optimizer.load_state_dict(torch.load(os.path.join(opt.start_from, 'optimizer.pth'))).
So does this mean that the learning_rate 5e-5 are discarded, and it's unnecessary for us to set this parameter when we train using self-critical?

[ruotianluo]

The learning rate will be updated according to the argument.

[siyilingting]

Thank you. But I am still puzzled which learning rate are used in self-critical, the 5e-5 or the one stored in optimizer.pth.

[ruotianluo]

I take it back. After reviewing my code, you are correct. This is definitely a bug. The learning rate will be what is saved in the optimizer.

[ruotianluo]

I fix it in the latest commit. Thank you for pointing this out.