train --year 2007 failed about deeplabv3plus-pytorch HOT 4 CLOSED

@ynjiun,

if you set the train loader to drop_last=True , you will not face this problem. 16*13=208, leaving one image behind, should not work for 16/13 but should had work for 13.

@VainF ,

I think it will be best if you set the train loader to drop_last=True just in case. Edit line: https://github.com/VainF/DeepLabV3Plus-Pytorch/blob/master/main.py#L236

Currently, pytorch batchnorm for training is not smart enough to ignore single batch size with single value feature map and some user will face this problem if their batch size has a remainder of 1 from the dataset.

For example, building a bigger backbone, I was forced to use batch_size=2 and on cityscapes, the error will pop up as the train set has 2975 images (leaving a remainder of 1 at the last load which causes the error).

from deeplabv3plus-pytorch.

Hi @ynjiun, please make sure that batch size>1 for training mode.

from deeplabv3plus-pytorch.

Hi VainF,

Thank you for your response. It turns out that for unknown reason if the batch_size is 16 or 15 or 13, it will report the error. But if I set batch_size 14 or 12, then it works. I don't know why. Any insights?

from deeplabv3plus-pytorch.

Fixed! Thanks for your help.

from deeplabv3plus-pytorch.