search

juhongm999 / hsnet

Official PyTorch Implementation of Hypercorrelation Squeeze for Few-Shot Segmentation, ICCV 2021
231 stars 43 forks source link

Question about learning rate #29

Closed DUT-CSJ closed 1 year ago

DUT-CSJ commented 1 year ago

Hello, thank you for sharing your work! When I try to reproduce hsnet, I find the learning rate is fixed through the training. I trained hsnet on pascal VOC2012 (1-shot, fold0) about 400 epochs, the miou on val set is very unstable (highest: 63.62 lower than 64.3 in paper). (gpu: single 3090, other settings are same to default.) Dose the lr need to be adjust? Looking forward to your reply.

juhongm999 commented 1 year ago

For accurate reproduction of our results, you should follow the same training settings as the ones we used. Specifically, our network was trained with a batch size of 20 and four Titan RTX GPUs over a couple of days. Please aim to replicate these settings as closely as possible.

DUT-CSJ commented 1 year ago

For accurate reproduction of our results, you should follow the same training settings as the ones we used. Specifically, our network was trained with a batch size of 20 and four Titan RTX GPUs over a couple of days. Please aim to replicate these settings as closely as possible.

I don't have the same gpu with yours, what else, thank you for your reply!