vyshnav112233/diabetic-retinopathy-efficientnet-b5

Diabetic Retinopathy EfficientNet-B5

This model is a timm EfficientNet-B5 trained as a single-output ordinal regression model for diabetic retinopathy severity grading.

Best local run directory: logs/run_20260510_154446

Training setup:

• —Model: EfficientNet-B5
• —Task: single-output regression over classes 0-4
• —Loss: sample-weighted MSE for training, plain MSE for validation/test
• —Optimizer: AdamW
• —Batch size: 16
• —Gradient accumulation: 4
• —Effective batch size: 64
• —Learning rate: 2e-4
• —Weight decay: 1e-4
• —Scheduler: 2-epoch linear warmup then cosine annealing to 1e-6
• —Tuned thresholds: [0.54, 1.62, 2.46, 3.36]

This checkpoint is for project/inference use only and is not a medical device.