Late submission attempt for Galaxy Zoo - The Galaxy Challenge on Kaggle.
This is my Final Year Project (Sem 1) PHYS4610 at CUHK. The report is provided in the repository.
Highlights:
- Constructed a CNN baseline model from scratch using the PyTorch framework
- Averaging multiple outputs by exploiting rotational and reflectional invariances greatly improve the score
- Baseline + CBAM (Convolutional Block Attention Module) reproduced single-model score of competition winner
- Transfer learning greatly improves the score and reduces training time
- Pretrained DINOv2 gave the best score
- Pretrained ResNet50 has the shortest runtime and a great performance
Best single-model score:
| Model | Score (RMSE) | # of Epochs | Pretrained dataset | Total runtime / per Epoch (s) |
|---|---|---|---|---|
| Baseline | 0.07742 | 240 | / | 19200 / 80 |
| Baseline + CBAM | 0.07693 | 240 | / | 19680 / 82 |
| ResNet50 (not pretrained) | 0.07574 | 120 | / | 15000 / 125 |
| ResNet50 | 0.07393 | 30 | ImageNet-1K v1 | 3750 / 125 |
| ResNet101 | 0.07400 | 30 | ImageNet-1K v1 | 5520 / 184 |
| ConvNeXt Large | 0.07357 | 30 | ImageNet-1K v1 | 45000 / 1500 |
| DINOv2 Base | 0.07251 | 15 | LVD-142M | 17400 / 580 |
