COMP6258 CoSNet Reproducibility Project

This repository is the project framework for the COMP6258 reproducibility challenge.

Paper

Designing Concise ConvNets with Columnar Stages, ICLR 2025.

Reproduction Scope

We plan to reproduce a restricted version of CoSNet on ImageNet-100.

Models:

• CoSNet-A0
• ResNet-18
• ConvNeXt-Tiny

Main metrics:

• Top-1 Accuracy
• Params
• FLOPs
• Latency / FPS
• Training and validation curves

Important Note

This repository is a skeleton framework only.

Most Python files contain only TODO placeholders. Each member should complete their own assigned part.

Team Responsibilities

Area	Owner	Files
CoSNet-A0	Member A	src/models/cosnet.py
ResNet-18 baseline	Member A	src/models/resnet.py, configs/resnet18.yaml
Training pipeline	Member A	src/train.py, src/evaluate.py, src/datasets.py, src/utils.py
ConvNeXt-Tiny baseline	Member B	src/models/convnext.py, configs/convnext_tiny.yaml
Params / FLOPs	Member C	src/complexity.py
Latency / FPS	Member C	src/latency.py
Plotting and figures	Member C	src/plot_results.py, src/cam.py
Report writing	All members	report/paper.tex

Expected Dataset Structure

data/imagenet100/
├── train/
│   ├── class_001/
│   ├── class_002/
│   └── ...
└── val/
    ├── class_001/
    ├── class_002/
    └── ...

Unified Log Format

All model training logs should use:

epoch,train_loss,val_loss,train_acc,val_acc,lr,time

Example:

epoch,train_loss,val_loss,train_acc,val_acc,lr,time
0,4.52,4.31,3.2,4.1,0.001,320
1,4.20,4.05,6.8,8.4,0.00099,318

Setup

python -m pip install -r requirements.txt

Planned Commands

Train CoSNet-A0:

python src/train.py --config configs/cosnet_a0.yaml

Train ResNet-18:

python src/train.py --config configs/resnet18.yaml

Train ConvNeXt-Tiny:

python src/train.py --config configs/convnext_tiny.yaml

Measure complexity:

python src/complexity.py --model cosnet_a0 --config configs/cosnet_a0.yaml

Measure latency:

python src/latency.py --model cosnet_a0 --config configs/cosnet_a0.yaml

Plot results:

python src/plot_results.py --input results/logs --output results/figures