search

juhongm999 / hsnet

Official PyTorch Implementation of Hypercorrelation Squeeze for Few-Shot Segmentation, ICCV 2021
231 stars 43 forks source link

Experiment without support feature masking #11

Closed AyanamiReiFan closed 3 years ago

AyanamiReiFan commented 3 years ago

How to do the experiment without support feature masking？Should I retrain whole model without support feature masking？

I simply delete the 52th line in hsnet.py during inference(use the checkpoint you given): support_feats = self.mask_feature(support_feats, support_mask.clone())

and got a poor result:

fold0
*** Test [@Epoch 00] Avg L: 0.00000  mIoU: 27.50   FB-IoU: 51.22   ***
fold1
*** Test [@Epoch 00] Avg L: 0.00000  mIoU: 41.96   FB-IoU: 58.30   ***
fold2
*** Test [@Epoch 00] Avg L: 0.00000  mIoU: 33.78   FB-IoU: 48.98   ***
fold3
*** Test [@Epoch 00] Avg L: 0.00000  mIoU: 32.07   FB-IoU: 51.38   ***
juhongm999 commented 3 years ago

Did you evaluate the provided models that are trained without support feature masking [link]?

AyanamiReiFan commented 3 years ago

Thanks，the result is right with this checkpoint