search

wds1998 / Edge-LBAM

Pytorch implementation of paper "Image Inpainting with Edge-guided Learnable Bidirectional Attention Maps"
23 stars 5 forks source link

Strange results #4

Open asdj32n12dk opened 3 years ago

asdj32n12dk commented 3 years ago

Hey, I'm getting strange results even on small masks. Could you confirm it should be like that or not?

[mask, damaged, generared, GT, outputs_merged]

Places365_val_00000079_10%-15%_14 73%_Places365_test_00180291 Places365_val_00000495_25%-30%_25 13%_Places365_test_00065180 Places365_val_00006189_00%-05%_04 18%_Places365_test_00273258 Places365_val_00007302_00%-05%_01 91%_Places365_test_00059890 Places365_val_00035912_00%-05%_00 37%_Places365_test_00270309 Places365_val_00032957_00%-05%_02 50%_Places365_test_00196871

BTW, the current test is still not working. So I had to make these changes: 1) fake_images = netG(inputImgs2, masks,outputs_merged) -> returns tuple, so I had to use g_image = fake_images[0].data.cpu() instead of g_image = fake_images.data.cpu() 2) The MECNet is not loading weights by default, so I set self.gen_weights_path = <path to MECNet.pth> manually in MECNet/models.py

delldu commented 2 years ago

Compare previous BLAM, this version is bad enough.