# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Repository overview

This repository currently contains the machine-learning pipeline for ZeaVis Edu: a corn leaf disease classifier trained with EfficientNetV2B0 and exported for production use as TensorFlow SavedModel, TFLite, and TensorFlow.js formats.

The active project lives under `Machine_Learning/`. Most commands should be run from that directory unless noted otherwise.

## Common commands

```bash
cd Machine_Learning
```

Set up a Python environment:

```bash
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
```

Run local dataset preprocessing after placing `dataset_1.zip`, `dataset_2.zip`, and `dataset_3.zip` beside `preprocessing.py`:

```bash
python preprocessing.py
```

Export a trained Keras model to SavedModel and TFLite after placing the Colab-trained model at `best_model/best_model.keras`:

```bash
python save_model.py
```

Convert the SavedModel export to TensorFlow.js via CLI:

```bash
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
tensorflowjs_converter \
  --input_format=tf_saved_model \
  --output_format=tfjs_graph_model \
  --signature_name=serving_default \
  --saved_model_tags=serve \
  model/saved_model \
  model/tfjs_model
```

Open the training notebook locally if needed:

```bash
jupyter notebook notebook.ipynb
```

There is no project test suite, lint command, or build system configured in the current repository.

## High-level architecture

- `Machine_Learning/preprocessing.py` prepares the training dataset locally. It extracts three source ZIP files, merges selected class folders into `dataset/`, maps selected Mandarin labels from Dataset 3 via `desc.json`, removes known problematic image files, then creates `dataset.zip` for upload to Google Drive/Colab.
- `Machine_Learning/notebook.ipynb` is the training workflow intended for Google Colab with GPU enabled. It trains an EfficientNetV2B0-based classifier and saves the best model to Google Drive as `best_model.keras`.
- `Machine_Learning/save_model.py` is the production export step. It loads `best_model/best_model.keras`, rebuilds a clean EfficientNetV2B0 architecture without training-time augmentation layers, copies weights into that model, exports `model/saved_model/`, and writes `model/model.tflite`.
- TensorFlow.js export is intentionally done with the `tensorflowjs_converter` CLI rather than from Python to avoid protobuf/runtime conflicts documented in the README.

## Model labels and dataset mapping

The classifier targets four Indonesian labels:

- `Bercak Daun` — Gray Leaf Spot
- `Hawar Daun` — Northern/Southern Leaf Blight
- `Karat Daun` — Common Rust
- `Daun Sehat` — healthy corn leaf

Dataset handling is part of the model logic:

- Dataset 1 contributes `Bercak Daun`, `Hawar Daun`, and `Daun Sehat`; its `Karat Daun` folder is intentionally ignored because the README states it is not representative.
- Dataset 2 contributes `Common_Rust` mapped to `Karat Daun` and `Healthy` mapped to `Daun Sehat`.
- Dataset 3 is routed through Mandarin label mappings in `PEMETAAN_KATEGORI` inside `preprocessing.py`.

## Important generated/local artifacts

The following files/directories are generated or externally supplied during the ML workflow and may not exist in a fresh clone:

- `Machine_Learning/dataset_1.zip`, `dataset_2.zip`, `dataset_3.zip` — manually downloaded source datasets.
- `Machine_Learning/dataset/` and `Machine_Learning/dataset.zip` — generated by `preprocessing.py`.
- `Machine_Learning/best_model/best_model.keras` — trained model downloaded from Colab/Google Drive.
- `Machine_Learning/model/saved_model/`, `model/model.tflite`, and `model/tfjs_model/` — production exports.

## Notes for future changes

- Keep README command examples and this file in sync when changing the ML pipeline.
- Preserve the current class label names unless the training notebook, preprocessing mappings, and downstream app/API expectations are updated together.
- `save_model.py` assumes the clean architecture matches the trained model weights exactly; changes to the notebook model architecture usually require corresponding changes in `build_clean_model()`.