Architecture: ResNet-50 (via timm, ImageNet pretrained)
Custom Head: Global Average Pool → BatchNorm → Dropout(0.4) → Linear(2048→256, ReLU) → Dropout(0.3) → Linear(256→5)
Input: 380×380 RGB fundus images
Output: 5-class probabilities
Parameters: 24M
Preprocessing: Ben Graham's method

Performance

Single train/val split (used as baseline):

Metric	Value
Best Val QWK	0.7281

Compared to EfficientNet-B4 primary model: ResNet-50 is ~0.06 QWK worse despite having more parameters (24M vs 18M).

Why include this baseline?

The ResNet-50 baseline establishes a reference for the EfficientNet-B4 improvement. It's intentionally trained with the same hyperparameters and pipeline so results are directly comparable.

Usage

Same as EfficientNet but use build_resnet:

python

from src.models.resnet_baseline import build_resnet
from src.utils.config import load_config
from src.utils.checkpoint import load_checkpoint
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(repo_id="anishanish383/dr-detection-resnet50", filename="best.pt")
config = load_config("configs/train_resnet.yaml")
model = build_resnet(config)
load_checkpoint(ckpt_path, model)
model.eval()

Limitations

Same as the EfficientNet model — this is a research baseline, not for clinical use. See the primary model card for full details.

License

MIT