agri-dataset-loader

Agricultural CV Dataset Loader

Description

Unified loading, preprocessing, and augmentation for agricultural computer vision datasets. Supports automatic download from Kaggle, HuggingFace Hub, and GitHub releases; deterministic train/val/test splitting; domain-specific augmentation presets; and cross-dataset label alignment for domain-adaptation experiments.

When to Use

• User mentions agricultural / crop / plant / disease related datasets
• Need to load PlantVillage, PlantDoc, DeepWeeds, or MinneApple
• Need to prepare train / val / test data for agricultural CV experiments
• Need domain-specific data augmentation for field imagery
• Need to align labels across PlantVillage (lab) and PlantDoc (field) for cross-domain transfer learning

Prerequisites

pip install torch torchvision albumentations pandas kaggle gdown Pillow tqdm

Kaggle API credentials (optional, for PlantVillage):

# ~/.kaggle/kaggle.json  (0600 permissions)
{"username": "YOUR_USERNAME", "key": "YOUR_KEY"}

Supported Datasets

Dataset	Task	Scale	Download
PlantVillage	Classification	54,305 imgs, 38 classes	Kaggle API / HuggingFace
PlantDoc	Classification, Detection	2,598 imgs, 27 classes	GitHub release
DeepWeeds	Classification	17,509 imgs, 9 classes	TensorFlow Datasets
MinneApple	Detection, Segmentation	1,000+ imgs	Direct URL
IP102	Classification	75,000 imgs, 102 classes	Application required

Quick Start

import torch
from agri_cv_research.datasets import PlantVillageDataset

# Auto-download and load
ds = PlantVillageDataset(
    root="./data/plantvillage",
    split="train",
    download=True,
    transform="default",
    target_size=(224, 224),
)

loader = torch.utils.data.DataLoader(
    ds, batch_size=32, shuffle=True, num_workers=4
)

for imgs, labels in loader:
    print(imgs.shape, labels.shape)  # torch.Size([32, 3, 224, 224])

Detailed Usage