search

hustvl / WeakTr

WeakTr: Exploring Plain Vision Transformer for Weakly-supervised Semantic Segmentation
MIT License
120 stars 2 forks source link

代码的OnlineRetrain是哪个版本 #15

Closed HAL-42 closed 1 year ago

HAL-42 commented 1 year ago

感谢你的工作！

这里想问一下，如果按照docs/Training.md 中方法做 Online Retraining，使用的是哪一个预训练模型？ 是pretrained on ImageNet-21k，能达到78.4% mIoU的那个模型吗？

谢谢！

Yingyue-L commented 1 year ago

You can reproduce the ImageNet-21k pretrained results(update to 78.8%) use the scripts below, please download the updated CAM label WeakTrV2_CAMlb_wCRF.

CUDA_VISIBLE_DEVICES=2 PORT=14888 bash segm/dist_train.sh 1 \
--log-dir start1.0_patch120_seg_vit_small_patch16_384_voc_weaktrv2 \
--dataset pascal_voc --backbone vit_small_patch16_384 --decoder mask_transformer \
--ann-dir WeakTrV2_CAMlb_wCRF \
--start-value 1.0 --patch-size 120 \
HAL-42 commented 1 year ago

有效 多谢！