pytorch-deeplab-xception
pytorch-deeplab-xception copied to clipboard
achieved 79.11% mIoU with Resnet101 backbone
First of all, thanks for creating this repo. It is very clear and easier to use than the original tensorflow model. I ran a few training experiments on VOC2012 and saw many people trying to train the network and achieve a good result as the uploaded checkpoints (78.43%). Like many I felt there is no clear answer for those of us who uses one gpu and have a memory problem using a batch size of 16. I wanted to share the script parameters to help other people in case they found it hard to train the network above 78%.
my mIoU maxed at 79.1% around epoch 45 with the following call: python train.py --backbone resnet --lr 0.0035 --workers 14 --epochs 50 --batch-size 8 --gpu-ids 0 --checkname deeplab-resnet --eval-interval 1 --dataset pascal
@Nadavc220 --workers 14, is it a typo?
Weirdly enough this is not a typo. I did not intend to run so many workers at all but I guess I did it by mistake when I ran the script. There is no reason to use 14 workers.
Yes I did. apart from the changes mentioned in the post everything was set as the default value.
On Sun, Jul 5, 2020 at 7:42 PM htwang14 [email protected] wrote:
Did you use SBD dataset to achieve this high mIoU?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/jfzhang95/pytorch-deeplab-xception/issues/172#issuecomment-653910605, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHCNQ3IVA2XDRCAH6AQQX7TR2CUN3ANCNFSM4NNKDAEA .
Thanks!
First of all, thanks for creating this repo. It is very clear and easier to use than the original tensorflow model. I ran a few training experiments on VOC2012 and saw many people trying to train the network and achieve a good result as the uploaded checkpoints (78.43%). Like many I felt there is no clear answer for those of us who uses one gpu and have a memory problem using a batch size of 16. I wanted to share the script parameters to help other people in case they found it hard to train the network above 78%.
my mIoU maxed at 79.1% around epoch 45 with the following call: python train.py --backbone resnet --lr 0.0035 --workers 14 --epochs 50 --batch-size 8 --gpu-ids 0 --checkname deeplab-resnet --eval-interval 1 --dataset pascal
excuse me, how to use the pascal 2012 dataset to train the model?