This codebase focus on "Explicitly Hierarchy" implemented in "Learning with Hierarchical Complement Objective" for anonymized code submission. Since the codebase of "Latent Hierarchy" involved many details of Semantic Segmentation, We will summarize it as soon as possible and open another repo to present.
- Python 3.6
- Pytorch 1.2.0
- keras
- tensorflow
- numpy 1.16.4
For getting baseline results
python main.py --sess Baseline_session
For training via Complement objective
python main.py --COT --sess COT_session
For training via Hierarchical Complement Entropy (HCE)
python main.py --HCOT --sess HCOT_session
The following table shows the test error rates in a 200-epoch training session. (Please refer to "Table 1" in the paper for details.)
Model | Baseline | COT | HCOT |
---|---|---|---|
PreAct ResNet-18 | 25.44% | 24.73% | 23.8% |