LikeLy-Journey/SegmenTron

Can't achieve the mIoU with your modelzoo

DePengW opened this issue · 2 comments

DePengW commented

I tried to train DeepLabv3_plus model for pascal_voc by using
CUDA_VISIBLE_DEVICES=0 python-u tools/train.py --config-file configs/pascak_voc_deeplabv3_plus.yaml
due to my configuration problem, modifying the batch size to 4, others is similar to you. But , to 50 epoch , I only achieve 62 mIoU.
Similarly, I tried to train DeepLabv3_plus model for cityscape by using
CUDA_VISIBLE_DEVICES=0 python -u tools/train.py --config-file configs/cityscapes_deeplabv3_plus.yaml
I also achieve about 65 mIoU. So , I want to know whether the pre-training model is used?

LikeLy-Journey commented

They all need pretrained models. pascal_voc_deeplabv3_plus.yaml need pretrained model on pascal_aug, cityscapes_deeplabv3_plus.yaml need pretrained model xception65. you can find this model in release. Also you shoud notice BATCH_SIZE in config is the image in a single gpu. you should set total batch size=16.

andrew-begain commented

They all need pretrained models. pascal_voc_deeplabv3_plus.yaml need pretrained model on pascal_aug, cityscapes_deeplabv3_plus.yaml need pretrained model xception65. you can find this model in release. Also you shoud notice BATCH_SIZE in config is the image in a single gpu. you should set total batch size=16.

@DePengW @LikeLy-Journey i train the pascal_voc_deeplabv3_plus.yaml ,and this need the pretrained model xception65,and i can not achieve the mIoU
image,LikeLy-Journey mentioned pascal_voc_deeplabv3_plus.yaml need pretrained model on pascal_aug,i want to know how to add the pretrained model on pascal_aug,thanks