fail to reproduce 79.19% IoU on val using train_aug

[Euphoria16]

Hi,great work and thanks for sharing your scripts.

May I ask a question? I cannot reproducing your 79.19% IoU although I didn't change the configs——I use 4 GPUs with batch size 16 each GPU(total 64) and I use train_aug(totally 10582 images). Is there something wrong? I got 75.01% highest IoU after 60 epochs.

Could you please specific your settings? Has anyone other reproduced the result successfully?

You said just run python train.py instead of using voc2012_train.sh, so does it mean you get 79.19% after only stage1 or after both stage1 and stage2? It seems you got this IoU only after 43 epochs so it is only after stage1 I guess. Is stage2 needed?

[zzl-997]

你好，想问一下这个ImageSets/Segmentation/train_aug.txt和ImageSets/Segmentation/val_aug.txt在哪，我在下载的数据集里没有找到

[CharlesPikachu]

you can try this repo https://github.com/SegmentationBLWX/sssegmentation with higher performance