Hi, I cannot reproduce the result.

irfanICMLL / structure_knowledge_distillation

The official code for the paper 'Structured Knowledge Distillation for Semantic Segmentation'. (CVPR 2019 ORAL) and extension to other tasks.

BSD 2-Clause "Simplified" License

708 stars 103 forks source link

Hi, I cannot reproduce the result. #22

Closed aachenhang closed 4 years ago

aachenhang commented 4 years ago

Hi，I have used your code and failed to reproduce the result. Here are my results.

I used the master branch. When load the imagenet weight, I skip 'conv1.weight' and 'layer1.0.conv1.weight' cause their shape is different with the official Resnet18. The results are all best IOU on the cityscape eval set during training.

Baseline: 0.619
Baseline + PI: 0.397
Baseline + PI + PA: 0.414
Baseline + PI + PA + HO: D_loss NAN

Could you send me your log?

aachenhang commented 4 years ago

By the way, I found that training more iteration on the baseline can make the result better.

Baseline: 0.619
- Resume the 0.619 ckpt and train more 40k iters without any distillation: 0.676
- Resume the 0.676 ckpt and train more 40k iters without any distillation: 0.700

irfanICMLL commented 4 years ago

I have already sent the training log to you through the email.

irfanICMLL commented 4 years ago

I will close the issue. If you have further questions, please feel free to contact me.

DanceInDark commented 4 years ago

Hi，I have used your code and failed to reproduce the result. Here are my results.

I used the master branch. When load the imagenet weight, I skip 'conv1.weight' and 'layer1.0.conv1.weight' cause their shape is different with the official Resnet18. The results are all best IOU on the cityscape eval set during training.

Baseline: 0.619

Baseline + PI: 0.397

Baseline + PI + PA: 0.414

Baseline + PI + PA + HO: D_loss NAN

Could you send me your log?

I also have the same problem and cannot reproduce the result. Have you fixed this?