search

HAL-42 / FMA-WSSS

This repo is a implementation of the Foundation Model Assisted Weakly Supervised Semantic Segmentation. The code is developed based on the Pytorch framework.
Apache License 2.0
40 stars 4 forks source link

In training classification task which is step 2 for the process. #33

Closed Shield9999 closed 3 months ago

Shield9999 commented 3 months ago

In training classification task which is step 2 for the process I get weight missed for the pascal and coco. Any solution for this? Or is this the right one.

image

HAL-42 commented 3 months ago

This shouldn't be a problem. The missing parameter is CLIP's pre-trained weights.

Shield9999 commented 3 months ago

Thx for your comment!

Shield9999 commented 2 months ago

To Yang Xiaobo

Hello author. I'm using your foundation model as my baseline to develop the performance. When I'm reproducing your baseline performance I could reproduce the COCO performance however for the pascal voc there are some gap between the experiment and your paper. 79.5 or 81.2 for the validation set, and the test set was fine. Could tell me more detail or why this performance gap is happening?

Best regards, David Jung(SHIELD 999)

2024년 7월 1일 (월) 오후 3:30, Yang Xiaobo @.***>님이 작성:

This shouldn't be a problem. The missing parameter is CLIP's pre-trained weights.

— Reply to this email directly, view it on GitHub https://github.com/HAL-42/FMA-WSSS/issues/33#issuecomment-2199342035, or unsubscribe https://github.com/notifications/unsubscribe-auth/A2HOTROYHW4YHT6IVD74YOTZKDZO3AVCNFSM6AAAAABKE4DMHWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOJZGM2DEMBTGU . You are receiving this because you authored the thread.Message ID: @.***>

HAL-42 commented 2 months ago

Here , all experimental results and checkpoints are provided. By comparing logs and results step by step, can you locate at which specific step the difference occurred?

Shield9999 commented 2 months ago

To. Yang Xiaobo

The experimental result on the step 'Train segmentation network with fine seed' part which command is bash tools/dist_train.sh configs/m2f_psudo/m2f-sl22-bt4-80k-512x-VOC.py 2'. At this step I used my seed and use author's provided seed too. However for me the best result was 81.3 for the validation step and for the author's seed the best mIoU was 81.73. However the best mIoU in the paper was 83.1. I want to reproduce this result. Could you give me some advice?

Best regards David Jung

2024년 7월 18일 (목) 오후 10:11, Yang Xiaobo @.***>님이 작성:

Here https://zjueducn-my.sharepoint.com/personal/hal_42_zju_edu_cn/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fhal%5F42%5Fzju%5Fedu%5Fcn%2FDocuments%2FTrans%2DWSSS%2DOS&ga=1 , all experimental results and checkpoints are provided. By comparing logs and results step by step, can you locate at which specific step the difference occurred?

— Reply to this email directly, view it on GitHub https://github.com/HAL-42/FMA-WSSS/issues/33#issuecomment-2236459722, or unsubscribe https://github.com/notifications/unsubscribe-auth/A2HOTRK3KKK3TVSNACV2QSTZM65IJAVCNFSM6AAAAABKE4DMHWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMZWGQ2TSNZSGI . You are receiving this because you modified the open/close state.Message ID: @.***>

HAL-42 commented 2 months ago

The best mIoU appeared at step=30000, not the final step. MMSeg should save a checkpoint named best_mIoU_iter_30000.pth. Diff your log with provied experiment\others\mmseg\m2f-sl22-bt4-80k-512x-VOC\20231208_223350\20231208_223350.log