import numpy as np, json
from huggingface_hub import hf_hub_download
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.mobilenet_v2 import preprocess_input

repo = "hirooshaweerasuriya/rice-leaf-disease-mobilenetv2"
model  = load_model(hf_hub_download(repo, "rice_disease_model.keras"))
config = json.load(open(hf_hub_download(repo, "config.json")))
labels = [config["id2label"][str(i)] for i in range(config["num_classes"])]

img = image.load_img("leaf.jpg", target_size=(224, 224))
x   = preprocess_input(np.expand_dims(image.img_to_array(img), 0))

probs = model.predict(x)[0]
print(labels[int(probs.argmax())], f"{probs.max():.1%}")

Preprocessing must match training: preprocess_input scales pixels to [-1, 1]. Using /255.0 instead will produce confident, wrong answers.

Training

Base: MobileNetV2, ImageNet weights, include_top=False
Head: GlobalAveragePooling2D → Dropout(0.3) → Dense(128, relu) → Dropout(0.2) → Dense(8, softmax)
Stage 1: backbone frozen, Adam lr=1e-3, 8 epochs
Stage 2: top 30 layers unfrozen (BatchNorm kept frozen), Adam lr=1e-5, 6 epochs
Augmentation: rotation 25°, zoom 0.2, shifts 0.1, shear 0.1, horizontal flip, brightness 0.8–1.2
Class imbalance handled with balanced class_weight
Data: anshulm257/rice-disease-dataset, 80/20 split
Trained on a Kaggle free-tier GPU notebook

Results

Metric	Value
Validation accuracy	0.7604
Validation loss	0.7156

Class	Val samples	Recall
Bacterial Leaf Blight	107	80.37%
Brown Spot	162	61.11%
Healthy Rice Leaf	102	91.18%
Leaf Blast	185	62.16%
Leaf scald	112	70.54%
Narrow Brown Leaf Spot	70	64.29%
Rice Hispa	132	87.88%
Sheath Blight	165	93.33%

Limitations

Trained on a single curated dataset; field photos with varied lighting, backgrounds and phone cameras will be harder.
The dataset appears pre-augmented, so near-duplicates may span the train/val split — real-world accuracy is likely below the number above.
Always returns one of the 8 classes; a photo of something else still gets a confident label. Threshold on max probability if that matters.
A decision-support tool, not a substitute for agronomist diagnosis.

Files

File	Description
`rice_disease_model.keras`	Keras 3 model
`saved_model/`	TensorFlow SavedModel
`model.tflite`	Quantised TFLite build for mobile
`config.json`	Labels, image size, preprocessing, metrics
`labels.txt`	Class names in index order