Segment labels mismatch
fungtion opened this issue · comments
The pretrained model ./checkpoints/epoch_0250_iter_050000.pth output 20 classes, but the out_channels in face_seg_single_view.yaml is 19, so it can not be loaded correctly. And in segfaces dataset, all segmaps share 19 classes:
[0 'background', 1 'skin', 2 'l_brow', 3 'r_brow', '4 l_eye', 5 'r_eye', 6' eye_g', 7 'l_ear', 8 'r_ear', 9 'ear_r', 10 'nose', 11 'mouth', 12 'u_lip', 13 'l_lip', 14 'neck', 15 'neck_l', 16 'cloth', 17 'hair', 18 'hat']
in CeleAMask dataset, all segmaps share 20 classes:
[0: 'background' 1: 'skin' 2: 'l_nose' 3: 'eye_g', 4: 'l_eye' 5: 'r_eye', 6: 'l_brow' 7: 'r_brow' 8: 'l_ear' 9: 'r_ear', 10: 'mouth' 11: 'u_lip' 12: 'l_lip' 13: 'hair' 14: 'hat' 15: 'ear_r' 16: 'neck_l' 17: 'neck' 18: 'cloth' 19: r_nose]
Are the output classes from the pretrained model same with CeleAMask?
Sorry, something wrong in the above, 20 classes in segfaces should be:
[0 'background', 1 'skin', 2 'l_brow', 3 'r_brow', '4 l_eye', 5 'r_eye', 6' l_nose', 7 'r_nose', 8 'mouth', 9 'u_lip', 10 'l_lip', 11 'l_ear', 12 'r_ear', 13 'ear_r', 14 'eye_g', 15 'neck', 16 'neck_l', 17 'cloth', 18 'hair', 19 'hat']
so we must remap before perform inverse training
Hi, I think we can directly use segNet-20Class.pth checkpoint, which predicts 20-class semantic masks.
@RaymondJiangkw You're right, thanks.