Skip to content

Latest commit

 

History

History
executable file
·
79 lines (72 loc) · 3.97 KB

various_feature_connections.md

File metadata and controls

executable file
·
79 lines (72 loc) · 3.97 KB

Running SEFD with various feature connections

In this repository, we provide training and testing codes for SEFD with various edge
Download the edge of weights and place it under ${ROOT}/feat1234 and ${ROOT}/feat234 ... .
If you want to check the pre-trained weight test, go straight to chapter "Test SEFD with various feature connections"

Training with various feature connections

❗We compared feature connections only with the Canny edge.❗
If you want to experiment with a different structure map, simply change the distillation_loss in ${ROOT}/yaml_training/3dpw_{your_structure_map}_distill.yml. ex) 'log_softmax_l1' => 'ATLoss'

  • Training-Code (basic feature connection is feat34)
CUDA_VISIBLE_DEVICES=0 python train.py --amp --continue --gpu 0 --cfg ../assets/yaml_training/3dpw_"{your_structure_map}"_distil_loss_"{choose_your_loss}".yml
  • example (change structure map & loss) If you use HED as a structure map and Loss as ATLoss, copy and paste ${ROOT}/yaml_training/3dpw_hed_distill.yml => ${ROOT}/yaml_training/3dpw_hed_distill_loss_AT.yml and line 17: distillation_loss: "ATLoss"
CUDA_VISIBLE_DEVICES=0 python train.py --amp --continue --gpu 0 --cfg ../assets/yaml_training/3dpw_hed_distil_loss_AT.yml
  • example (change feature connections log softmax l1 loss) If you change feature connection; ex) feat34 => feat 234 check it ${ROOT}/main/model.py
    model.py => line in 554~566
#b_feature_1=F.log_softmax(b_feature_1,dim=1)
#feature1=F.log_softmax(feature1,dim=1)
#b_feature_2=F.log_softmax(b_feature_2,dim=1)
#feature2=F.log_softmax(feature2,dim=1)
b_feature_3=F.log_softmax(b_feature_3,dim=1)
feature3=F.log_softmax(feature3,dim=1)
b_feature_4=F.log_softmax(pose_guided_img_feat,dim=1)
feature4=F.log_softmax(feature4,dim=1)
#loss['kd_feature_loss_1'] = self.l1_loss(b_feature_1,feature1)
#loss['kd_feature_loss_2'] = self.l1_loss(b_feature_2,feature2)
loss['logsoftmax_feature_3'] = self.l1_loss(b_feature_3,feature3)
loss['logsoftmax_feature_4'] = self.l1_loss(b_feature_4,feature4)

You can get rid of the # you want. feat34 => feat234

#b_feature_1=F.log_softmax(b_feature_1,dim=1)
#feature1=F.log_softmax(feature1,dim=1)
b_feature_2=F.log_softmax(b_feature_2,dim=1)
feature2=F.log_softmax(feature2,dim=1)
b_feature_3=F.log_softmax(b_feature_3,dim=1)
feature3=F.log_softmax(feature3,dim=1)
b_feature_4=F.log_softmax(pose_guided_img_feat,dim=1)
feature4=F.log_softmax(feature4,dim=1)
#loss['kd_feature_loss_1'] = self.l1_loss(b_feature_1,feature1)
loss['kd_feature_loss_2'] = self.l1_loss(b_feature_2,feature2)
loss['logsoftmax_feature_3'] = self.l1_loss(b_feature_3,feature3)
loss['logsoftmax_feature_4'] = self.l1_loss(b_feature_4,feature4)
  • example (change feature connections ATloss) If you change feature connection; ex) feat1234 => feat 234 check it ${ROOT}/main/model.py
    model.py => line in 551~553
elif cfg.distillation_loss=="ATLoss":
    atl=self.atloss([b_feature_1,b_feature_2,b_feature_3,pose_guided_img_feat],[feature1,feature2,feature3,feature4])
    loss['ATLoss']=atl

You can get rid of the pose_guided_img_feat & feature4 you want. feat1234 => feat234

elif cfg.distillation_loss=="ATLoss":
    atl=self.atloss([b_feature_1,b_feature_2,b_feature_3],[feature1,feature2,feature3])
    loss['ATLoss']=atl
  • 👍 Other losses can be deleted in the same framework. 👍

Test SEFD with various feature connections

  • Pre-trained model testing (Only Canny)
CUDA_VISIBLE_DEVICES=0 python test.py --amp --continue --gpu 0 --cfg ../assets/yaml_test/3dpw_canny_distil_test.yml --exp_dir ../"{feature_connection}" --test_epoch 10
  • model testing (You should use the structural map you trained.)
CUDA_VISIBLE_DEVICES=0 python test.py --amp --continue --gpu 0 --cfg ../assets/yaml_test/3dpw_"{your_strcuture_map}"_distil_test.yml --exp_dir ../output/"{your_Exp_dir}" --test_epoch "your_epoch"