search

jiwoon-ahn / psa

Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation, CVPR 2018
MIT License
380 stars 62 forks source link

The training params for ResNet38_aff #18

Closed YudeWang closed 4 years ago

YudeWang commented 5 years ago

@jiwoon-ahn hi, Thanks for your nice work! I load your shared weights and achieve the same mIoU as you showed in README.

While I meet some trouble to train the model by myself and only achieve 59.077% mIoU (instead of 60.2%) on val set by ResNet38_aff. As you mentioned in other issues that the default params setting is just for vgg16, could you share the params setting for resnet38_aff?

Here is my res38_aff training setttings: lr=0.01 gpu=4 (may cause different batch number on each gpu) batchsize=8 max_epoch=8 loss_weight=1/4 bg + 1/4 fg + 1/2 * neg pretrained_model=res38_cls.pth (provided by you) alpha=4/16/32 other default params...

Seohyeong commented 4 years ago

I'm having difficulty training AffinityNet (corresponding to step 3. Train AffinityNet with the labels and 4. Perform Random Walks on CAMs), getting mIoU below 50%. Could you tell me which weights you used getting 59.077% mIoU on validation set? Did you use res38_cls.pth(output of step1) to do step 3 and used produced resnet38_aff (output of step 3) as weights for step 4?

YudeWang commented 4 years ago

@SeoHyeong I remember I use imagenet pretrained resnet38 weights and I achieve the performance in his README table with 2 gpus instead of 4.

Seohyeong commented 4 years ago

@YudeWang Thanks for your reply! I'll give it a try again with the pretrained weights.

yuleung commented 3 years ago

@YudeWang I have a question that what does the alpha=4/16/32 mean?
In ResNet38_aff, alpha was be set 8? I'm confused about it. So, Could you please tell me the reason？ Very grateful to you！