The loss can't be reduced when I use VOC dataset

[JoyHuYY1412]

I tried to use this code for the segmentation of VOC dataset. It has about 2900 images, I use the deeplabV2 model.
The loss hardly decreases after 2k steps (I use 3 as batch size ), and it's around 0.7-1.
I don't know why.

[kazuto1011]

Did you switch all the batch norm to eval mode? I received the report the model could be trained on Pascal VOC 2012, although the result was 1 point lower (here). Basically, I think you only have to replace the COCO-Stuff data loader with a third-party one for VOC. Anyway, please try with a default setting (10 samples as a batch, 20k steps)

[JoyHuYY1412]

Thank you for your reply. It works really well. I have another question, the init model, deeplabv2_resnet101_VOC2012.pth, is it trained on resnet for image classification or something?

[kazuto1011]

An init model is actually deeplabv2_resnet101_COCO_init.pth, which is "ResNet-101 + ASPP" model trained on MSCOCO segmentation set. The "ResNet" part was pre-trained on ImageNet.
The deeplabv2_resnet101_VOC2012.pth that you mentioned is a model which was initialized with MSCOCO parameters above and trained on Pascal VOC 2012.