Training Curves

scripts/plot_training.py visualizes the training history saved by train.py in training_history.json.

Usage

# Default (reads checkpoints/training_history.json)
python scripts/plot_training.py

# Custom paths
python scripts/plot_training.py --history checkpoints/training_history.json --outdir outputs/plots

CLI Reference

Flag	Default	Description
--history	checkpoints/training_history.json	Path to history JSON
--outdir	outputs/plots	Output directory for the PNG

Output

A single PNG (training_curves.png) with six panels (four if metrics are not yet available):

Panel	Content
Total Loss	Train + val total loss per epoch
Species Loss	Train + val species loss
Environmental Loss	Train + val env MSE
Learning Rate	LR schedule (log scale)
Validation mAP	Mean average precision (if recorded)
Top-k Recall	Top-10 and top-30 recall (if recorded)

Note

Metric panels (mAP, top-k recall) are only shown when the training history contains val_map and val_top10_recall keys. These are recorded automatically by train.py starting with the current version.