iLearn-Lab/ICMR25-HMKD

🚀 HMKD-ICMR: Heterogeneous Model Knowledge Distillation via Dual Alignment for Semantic Segmentation

Mingzhu Xu¹  Jing Wang¹  Mingcai Wang¹  Yiping Li¹  Yupeng Hu^1✉  Xuemeng Song¹  Weili Guan¹

¹Affiliation (Please update if needed)

Official implementation of HMKD, a Heterogeneous Model Knowledge Distillation framework with Dual Alignment for Semantic Segmentation.

🔗 Conference: ICMR 2025
🔗 Task: Semantic Segmentation
🔗 Framework: PyTorch

---

📌 Model Information

1. Model Name

HMKD (Heterogeneous Model Knowledge Distillation)

---

2. Task Type & Applicable Tasks

• Task Type: Semantic Segmentation / Model Compression
• Core Task: Knowledge Distillation for segmentation
• Applicable Scenarios:
  • Lightweight model deployment
  • Cross-architecture distillation
  • Efficient semantic understanding

---

3. Project Introduction

Semantic segmentation models often rely on heavy architectures, limiting their deployment in resource-constrained environments. Knowledge distillation (KD) provides a promising solution by transferring knowledge from a large teacher model to a compact student model.

HMKD introduces a Dual Alignment Distillation Framework, which:

• Aligns heterogeneous architectures between teacher and student models
• Performs feature-level and prediction-level alignment
• Bridges the representation gap across different model families
• Improves segmentation accuracy while maintaining efficiency

---

4. Training Data Source

Supported datasets:

• Cityscapes
• CamVid

Dataset	Train	Val	Test	Classes
Cityscapes	2975	500	1525	19
CamVid	367	101	233	11

---

🚀 Environment Setup

• Ubuntu 20.04.4 LTS
• Python 3.8.10 (Anaconda recommended)
• CUDA 11.3
• PyTorch 1.11.0
• NCCL 2.10.3

Install dependencies:

pip install timm==0.3.2
pip install mmcv-full==1.2.7
pip install opencv-python==4.5.1.48

---

⚙️ Pre-trained Weights

Initialization Weights

• ResNet-18
• ResNet-101
• SegFormer-B0
• SegFormer-B4

(Download from official PyTorch and Google Drive links)

---

Trained Weights

Download trained HMKD models:

• Baidu Cloud: https://pan.baidu.com/s/1xw_6ts5VNV73vXeOLAokwQ?pwd=jvx8

---

🚀 Training

1. Download datasets and pre-trained weights
2. Generate dataset path lists (.txt files)
3. Update dataset paths in the code

Run training:

CUDA_VISIBLE_DEVICES=0,1 nohup python -m torch.distributed.launch --nproc_per_node=2 train_NEW_AEU_kd.py > train.log 2>&1 &

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train_NEW_AEU_kd.py

---

⚠️ Notes

• Designed for research purposes
• Performance depends on teacher-student architecture pairing
• Multi-GPU training is recommended

---

📝 Citation

@ARTICLE{HMKD,
  author={Xu, Mingzhu and Wang, Jing and Wang, Mingcai and Li, Yiping and Hu, Yupeng and Song, Xuemeng and Guan, Weili},
  journal={ICMR}, 
  title={Heterogeneous Model Knowledge Distillation via Dual Alignment for Semantic Segmentation}, 
  year={2025}
}

---

📬 Contact

For questions or collaboration, please contact the corresponding author.

---