Reproduce issue. about deeplabv3plus-pytorch HOT 5 CLOSED

Hi @leoshine , which dataset was used for your training? voc12 or voc12aug? The pretrained models are trained on voc12aug which has more training data than the standard voc12.

from deeplabv3plus-pytorch.

Yes. I use voc12aug. I just turn off the visualization.
Nothing is changed:

python main.py --model deeplabv3plus_resnet101 --gpu_id 0,1,2,3 --year 2012_aug --crop_val --lr 0.01 --crop_size 513 --batch_size 16 --output_stride 16

B.t.w. I downloaded the released model (DeepLabV3Plus-ResNet101) and tested it. The mIOU can exactly reach 0.783.

from deeplabv3plus-pytorch.

#5 (comment).
This issue is caused by the batchnorm. There is no sync BN in this repo and training with multiple GPUs will simply split the batch and distribute them to each device. So each GPU only gets 16 / 4 = 4 samples and there is no communication between different devices. It is recommended to train the model on a single GPU or use some third-party SyncBN to replace the default one.

May be you can try the SyncBN from Hang Zhang.

from deeplabv3plus-pytorch.

Thanks for the information.
I've tried this version of SyncBN: https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
And I can now reproduce the mIOU=78.3%.
Great work. Thanks.

from deeplabv3plus-pytorch.

Yes. I use voc12aug. I just turn off the visualization.
Nothing is changed:

python main.py --model deeplabv3plus_resnet101 --gpu_id 0,1,2,3 --year 2012_aug --crop_val --lr 0.01 --crop_size 513 --batch_size 16 --output_stride 16

B.t.w. I downloaded the released model (DeepLabV3Plus-ResNet101) and tested it. The mIOU can exactly reach 0.783.

How do you test and get 0.783. Are there any tricks when testing since I can not get this point with the released model.

from deeplabv3plus-pytorch.