cfg.TRAIN.COMPUTE_PRECISE_BN about video-nonlocal-net HOT 4 CLOSED

Hi,
PRECISE_BN deos not affect the performance in the final model really, it just makes the train/val error curve looks better when you visualize it.

I have also tried using 16 samples per batch, but i actually did not get much improvement in the case of training with 400K iterations.

from video-nonlocal-net.

Thank you for your replay, when I use 16 samples per batch, I can get the performacne in the paper, but 8 samples per batch is 1.5% than 16 samples per batch.

from video-nonlocal-net.

Hi,
In that case, i am actually suspecting your lr schedule might not be well adjusted. If you are using a 4GPUs machine and running with our 8GPU script, remember to adjust the lr by using lr = lr/2 and double the training iterations, with batchsize=8 samples per GPU.

from video-nonlocal-net.

Hi,
ALL training use the same configs(base_lr=0.01 and total training epochs=90), except the batch_size per GPU(8 or 16).

Because I use my own code and pre_train model(not your release code and pre_train model), the config may be not suitable for me? I have try some experiments, expect have a good result.

from video-nonlocal-net.