plant-leaf-masterdataset-resnet18

Fine-tuned microsoft/resnet-18 on the Plant Leaf Disease MasterDataset (86 classes, 19 crops). Live pin: tag v1.0.1.

Test metrics (held-out, scored once)

Serving-aligned processor (crop_pct=1.0, bilinear, 224 short-side resize + 224 centre crop), temperature 1.088 applied. n = 14,454.

The test split was also viewed once during diagnosis of a Trainer eval-cache bug (dataloader_persistent_workers reused the validation loader). These numbers come from a standalone scorer (model/score_split.py) after that fix and after aligning the processor.

Validation (for calibration only)

n = 14,356. Temperature 1.0877, confidence threshold 0.30 (min precision 0.70, max pass rate). Accuracy on the serving path after alignment: 0.9614.

Preprocessing

Eval / serving must match training eval transforms:

• Resize shortest edge to 224 (bilinear)
• Centre crop 224
• ImageNet mean/std

v1.0 shipped a ConvNeXT processor with crop_pct=0.875 (resize to 256). v1.0.1 sets crop_pct=1.0 and bilinear resampling; weights are unchanged.

PlantDoc out-of-dataset

21 overlapping classes. 8 PlantDoc near-duplicates of MasterDataset train images were excluded from the OOD set (the train copies stay). Evaluated on 1,988 images with zero remaining pHash hits at distance ≤ 3.

Training

5 epochs, batch 64, lr 5e-4, cosine, warmup 500, seed 42. Best checkpoint by validation macro-F1. PlantDoc was not mixed into training.