You can deploy this model and start earning money today!

swaggerlish/
farmguard-multi-crops-disease-screen

Name: farmguard-multi-crops-disease-screen
Availability: InStock

Views

No views yet

FarmGuard AI Model README

This document describes the trained crop-disease classifier model, how to reproduce training/evaluation, and how to publish model artifacts to Hugging Face.

Model Summary

Task: Multi-class image classification of crop leaf diseases.
Architecture: efficientnet_b0 with a replaced classifier head.
Framework: PyTorch + torchvision.
Classes: 18 total classes across cassava, maize, pepper, and tomato labels.

Training Pipeline

Training is implemented in src/train.py:

Stage 1: train classifier head with frozen backbone.
Stage 2: unfreeze top feature blocks and fine-tune.
Best model selected by validation macro-F1 and saved as checkpoint.

Default output artifacts:

outputs/checkpoints/best_model.pth
outputs/checkpoints/class_to_idx.json

Dataset Choice and Access Rationale

PlantVillage (Kaggle dataset: abdallahalidev/plantvillage-dataset) is used for tomato/maize/pepper/corn classes.
Cassava Disease (Kaggle competition data, manually downloaded) is used for cassava classes.

Why cassava was downloaded manually

Automated cassava download in notebook/runtime can fail when competition access is restricted by authentication/rules acceptance in the execution environment. To keep training reproducible, cassava files were downloaded manually from Kaggle and then uploaded to the runtime before preparation.

Expected manual cassava layout

After upload/extraction, cassava data should be placed under data/raw in one of these supported formats:

train.csv + train_images/
class folders (cbb, cbsd, cgm, cmd, healthy) or equivalent verbose names handled by alias mapping in prepare_dataset.py.

Dataset Preparation

Dataset preparation is implemented in src/prepare_dataset.py:

PlantVillage source via Kaggle dataset download (abdallahalidev/plantvillage-dataset).
Cassava source is manually downloaded Kaggle competition data uploaded into data/raw (primary workflow).
For custom folder datasets, preferred class folder names are: cbb, cbsd, cgm, cmd, healthy.
Optional: set CASSAVA_KAGGLEHUB_DATASET only if you want non-competition cassava auto-download.
Robust folder-name normalization/mapping for variant class names.
Processed split output to:
- data/processed/train
- data/processed/val
- data/processed/test

Evaluate the Model

Evaluation CLI is implemented in src/evaluate.py.

Run:

bash

python -m src.evaluate \
  --checkpoint outputs/checkpoints/best_model.pth \
  --class-map outputs/checkpoints/class_to_idx.json

Generated evaluation artifacts:

outputs/checkpoints/test_classification_report.txt
outputs/checkpoints/test_confusion_matrix.json

Run Inference on One Image

Inference CLI is implemented in src/infer.py.

bash

python -m src.infer \
  --image /path/to/leaf.jpg \
  --checkpoint outputs/checkpoints/best_model.pth \
  --class-map outputs/checkpoints/class_to_idx.json \
  --top-k 3

End-to-End Colab Commands

From ml/ directory:

bash

pip install -r requirements.txt
# optional ONLY if you want non-competition cassava auto-download
# export CASSAVA_KAGGLEHUB_DATASET="<owner/dataset-slug>"
python -m src.prepare_dataset
python -m src.train
python -m src.evaluate

Publish to Hugging Face

Required package is already listed in requirements.txt (huggingface_hub).

Example upload script:

python

import os
from pathlib import Path
from huggingface_hub import HfApi

repo_id = "your-hf-username/farmguard-ai-crop-disease-model"
token = os.environ["HF_TOKEN"]

api = HfApi()
api.create_repo(repo_id=repo_id, repo_type="model", exist_ok=True, token=token)

for artifact in [
    "outputs/checkpoints/best_model.pth",
    "outputs/checkpoints/class_to_idx.json",
    "outputs/checkpoints/test_classification_report.txt",
    "outputs/checkpoints/test_confusion_matrix.json",
]:
    p = Path(artifact)
    if p.exists():
        api.upload_file(
            path_or_fileobj=str(p),
            path_in_repo=p.name,
            repo_id=repo_id,
            repo_type="model",
            token=token,
        )

Notes

train_config.yaml exists, but current training code uses hardcoded defaults in src/train.py.
For reproducibility, consider pinning dependency versions in requirements.txt.

swaggerlish/farmguard-multi-crops-disease-screen

FarmGuard AI Model README

Model Summary

Training Pipeline

Dataset Choice and Access Rationale

Why cassava was downloaded manually

Expected manual cassava layout

Dataset Preparation

Evaluate the Model

Run Inference on One Image

End-to-End Colab Commands

Publish to Hugging Face

Notes

swaggerlish/
farmguard-multi-crops-disease-screen