Thank you for your nice experiment.
I have a question in train.
In the paper, you say that you ran 200k iterations on batch number 16, but when I look at GoPro-AdaRevIDB-pretrain-4gpu.yml, it says as a comment.
Split 300k iterations into two cycles.
1st cycle: fixed 3e-4 LR for 92k iters.
2nd cycle: cosine annealing (3e-4 to 1e-6) for 208k iters.
I'm wondering how exactly it was trained since it looks like this
And I wonder if the encoder is learning by freezing the weights from start to finish?
Thank you!
Thank you for your nice experiment.
I have a question in train.
In the paper, you say that you ran 200k iterations on batch number 16, but when I look at GoPro-AdaRevIDB-pretrain-4gpu.yml, it says as a comment.
Split 300k iterations into two cycles.
1st cycle: fixed 3e-4 LR for 92k iters.
2nd cycle: cosine annealing (3e-4 to 1e-6) for 208k iters.
I'm wondering how exactly it was trained since it looks like this
And I wonder if the encoder is learning by freezing the weights from start to finish?
Thank you!