cassava-leaf-disease-efficientnet-b0

Cassava Leaf Disease Classifier (EfficientNet-B0) — v2

Fine-tuned EfficientNet-B0 classifying cassava leaf photos into 5 categories, trained on the Cassava Leaf Disease Classification dataset (~21,400 real field images from Uganda, Makerere AI Lab).

First step toward offline, edge-deployable computer vision for physical-world problems in Africa.

Classes

• 0: Cassava Bacterial Blight (CBB)
• 1: Cassava Brown Streak Disease (CBSD)
• 2: Cassava Green Mottle (CGM)
• 3: Cassava Mosaic Disease (CMD)
• 4: Healthy

Results (held-out 20% stratified validation)

• Accuracy: 0.800
• Macro F1: 0.684
• Best class: Mosaic (F1 0.911). Hardest: Bacterial Blight (F1 0.531).

Training

• Base: EfficientNet-B0 (ImageNet pretrained, timm)
• 224x224, ImageNet normalization; augmentation: flips, rotation, color jitter
• Class-weighted CrossEntropy (inverse-frequency) for heavy imbalance (~61% Mosaic)
• AdamW, lr 1e-4, CosineAnnealing over 12 epochs, best-epoch by macro-F1
• Hardware: Kaggle Tesla T4

Known limitation (important)

The model sometimes classifies mild or early-stage diseased leaves as Healthy — the highest-stakes error for real deployment, since it misses infections. This is worst for early Bacterial Blight. Faint disease signal on mostly-green leaves is the model's core remaining weakness and is not fully solved by this version.

Usage

import timm, torch
model = timm.create_model("efficientnet_b0", pretrained=False, num_classes=5)
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
model.eval()